It's finally finished!!!!!!! The reason the Attributes API was kept
private in 4.2 was due to some publicly visible implementation details.
It was previously implemented by overloading `columns` and
`columns_hash`, to make them return column objects which were modified
with the attribute information.

This meant that those methods LIED! We didn't change the database
schema. We changed the attribute information on the class. That is
wrong! It should be the other way around, where schema loading just
calls the attributes API for you. And now it does!

Yes, this means that there is nothing that happens in automatic schema
loading that you couldn't manually do yourself. (There's still some
funky cases where we hit the connection adapter that I need to handle,
before we can turn off automatic schema detection entirely.)

There were a few weird test failures caused by this that had to be
fixed. The main source came from the fact that the attribute methods are
now defined in terms of `attribute_names`, which has a clause like
`return [] unless table_exists?`. I don't *think* this is an issue,
since the only place this caused failures were in a fake adapter which
didn't override `table_exists?`.

Additionally, there were a few cases where tests were failing because a
migration was run, but the model was not reloaded. I'm not sure why
these started failing from this change, I might need to clear an
additional cache in `reload_schema_from_cache`. Again, since this is not
normal usage, and it's expected that `reset_column_information` will be
called after the table is modified, I don't think it's a problem.

Still, test failures that were unrelated to the change are worrying, and
I need to dig into them further.

Finally, I spent a lot of time debugging issues with the mutex used in
`define_attribute_methods`. I think we can just remove that method
entirely, and define the attribute methods *manually* in the call to
`define_attribute`, which would simplify the code *tremendously*.

Ok. now to make this damn thing public, and work on moving it up to
Active Model.

This method doesn't need to be lazy, as it is never called from reads.
The only time it is called are in write cases, where we're about to loop
through the results of it, and build the attribute objects anyway. So we
don't gain anything by dodging the instantiation here. This is the only
method that coupled `AttributeSet` to `LazyAttributeHash`, so removing
it puts us back in a place where we can use a normal hash instead.

Remaining are `limit`, `precision`, `scale`, and `type` (the symbol
version). These will remain on the column, since they mirror the options
to the `column` method in the schema definition DSL

The goal is to remove the type object from the column, and remove
columns from the type casting process entirely. The primary motivation
for this is clarity. The connection adapter does not have sufficient
type information, since the type we want to work with might have been
overriden at the class level. By taking this object from the column,
it is easy to mistakenly think that the column object which exists on
the connection adapter is sufficient. It isn't.

A concrete example of this is `serialize`. In 4.2 and earlier, `where`
worked in a very inconsistent and confusing manner. If you passed a
single value to `where`, it would serialize it before querying, and do
the right thing. However, passing it as part of an array, hash, or range
would cause it to not work. This is because it would stop using prepared
statements, so the type casting would come from arel. Arel would have no
choice but to get the column from the connection adapter, which would
treat it as any other string column, and query for the wrong value.

There are a handful of cases where using the column object to find the
cast type is appropriate. These are cases where there is not actually a
class involved, such as the migration DSL, or fixtures. For all other
cases, the API should be designed as such that the type is provided
before we get to the connection adapter. (For an example of this, see
the work done to decorate the arel table object with a type caster, or
the introduction of `QueryAttribute` to `Relation`).

There are times that it is appropriate to use information from the
column to change behavior in the connection adapter. These cases are
when the primitive used to represent that type before it goes to the
database does not sufficiently express what needs to happen. An example
of this that affects every adapter is binary vs varchar, where the
primitive used for both is a string. In this case it is appropriate to
look at the column object to determine which quoting method to use, as
this is something schema dependent.

An example of something which would not be appropriate is to look at the
type and see that it is a datetime, and performing string parsing when
given a string instead of a date.  This is the type of logic that should
live entirely on the type. The value which comes out of the type should
be a sufficiently generic primitive that the adapter can be expected to
know how to work with it.

The one place that is still using the column for type information which
should not be necessary is the connection adapter type caster which is
sometimes given to the arel table when we can't find the associated
table. This will hopefully go away in the near future.

Collection associations would have already been validated, but singular
associations were not.

Fixes #18735.

Fixes #18717

All columns which would map to a string primitive need this behavior.
Binary has it's own marker type, so it won't go through this conversion.
String and text, which need this, will.

Fixes #18585.

    Post.where('id = 1').or(Post.where('id = 2'))
    # => SELECT * FROM posts WHERE (id = 1) OR (id = 2)

[Matthew Draper & Gael Muller]

These callbacks will already have been defined when the association was
built. The check against `reflection.autosave` happens at call time, not
at define time, so simply modifying the reflection is sufficient.

Fixes #18704

Fixes #18696.

`bound_attributes` is now used universally across the board, removing
the need for the conversion layer. These changes are mostly mechanical,
with the exception of the log subscriber. Additional, we had to
implement `hash` on the attribute objects, so they could be used as a
key for query caching.

The column is primarily used for type casting, which we're trying to
separate from the idea of a column. Since what we really need is the
combination of a name, type, and value, let's use the object that we
already have to represent that concept, rather than this tuple. No
consumers of the bind values have been changed, only the producers
(outside of tests which care too much about internals). This is
*finally* possible since the bind values are now produced from a
reasonable number of lcoations.

I'm going to be extracting this logic into a clause class, things need
to go through a method and not access the values hash directly.

Attempting to grok this code by refactoring it as I go through it.

The only place it was accessed was in tests. Many of them have another
way that they can test their behavior, that doesn't involve reaching
into internals as far as they did. `AssociationScopeTest` is testing a
situation where the where clause would have one bind param per
predicate, so it can just ignore the predicates entirely. The where
chain test was primarly duplicating the logic tested on `WhereClause`
directly, so I instead just make sure it calls the appropriate method
which is fully tested in isolation.

The bind values can come from four places. `having`, `where`, `joins`,
and `from` when selecting from a subquery that contains binds. These
need to be kept in a specific order, since the clauses will always
appear in that order. Up until recently, they were not.

Additionally, `joins` actually did keep its bind values in a separate
location (presumably because it's the only case that people noticed was
broken). However, this meant that anything accessing just `bind_values`
was broken (which most places were). This is no longer possible, there
is only a single way to access the bind values, and it includes joins in
the proper location. The setter was removed yesterday, so breaking `+=`
cases is not possible.

I'm still not happy that `joins` is putting it's bind values on the
Arel AST, and I'm planning on refactoring it further, but this removes a
ton of bug cases.

3729103e

[ci skip]

Overriding these methods may cause unexpected results since
"table_name=" does more then just setting the "@table_name".

ref: https://github.com/rails/rails/pull/18622#issuecomment-70874358

Contrary to my previous commit message, it wasn't overkill, and led to
much cleaner code.

[Sean Griffin & anthonynavarre]

The last place that was assigning it was when `from` is called with a
relation to use as a subquery. The implementation was actually
completely broken, and would break if you called `from` more than once,
or if you called it on a relation, which also had its own join clause,
as the bind values would get completely scrambled. The simplest solution
was to just move it into its own array, since creating a `FromClause`
class for this would be overkill.

All of its uses have been moved to better places

This removes the need to duplicate much of the logic in `WhereClause`
and `PredicateBuilder`, simplifies the code, removes the need for the
connection adapter to be continuously passed around, and removes one
place that cares about the internal representation of `bind_values`

Part of the larger refactoring to change how binds are represented
internally

[Sean Griffin & anthonynavarre]

The Relation will ultimately end up holding a reference to the arel
table object, and its associated type caster. If this is a
`TypeCaster::Connection`, that means it'll hold a reference to the
connection adapter, which cannot be marshalled. We can work around this
by just holding onto the class object instead. It's ugly, but I'm hoping
to remove the need for the connection adapter type caster in the future
anyway.

[Sean Griffin & anthonynavarre]

This fixed an issue where `having` can only be called after the last
call to `where`, because it messes with the same `bind_values` array.
With this change, the two can be called as many times as needed, in any
order, and the final query will be correct. However, once something
assigns `bind_values`, that stops. This is because we have to move all
of the bind values from the having clause over to the where clause since
we can't differentiate the two, and assignment was likely in the form
of:

`relation.bind_values += other.bind_values`

This will go away once we remove all places that are assigning
`bind_values`, which is next on the list.

While this fixes a bug that was present in at least 4.2 (more likely
present going back as far as 3.0, becoming more likely in 4.1 and later
as we switched to prepared statements in more cases), I don't think this
can be easily backported. The internal changes to `Relation` are
non-trivial, anything that involves modifying the `bind_values` array
would need to change, and I'm not confident that we have sufficient test
coverage of all of those locations (when `having` was called with a hash
that could generate bind values).

[Sean Griffin & anthonynavarre]

When we made sure that the counter gets updated in memory, we only did
it on the has many side. The has many side only does the update if the
belongs to cannot. The belongs to side was updated to update the counter
cache (if it is able). This means that we need to check if the
belongs_to is able to update in memory on the has_many side.

We also found an inconsistency where the reflection names were used to
grab the association which should update the counter cache. Since
reflection names are now strings, this means it was using a different
instance than the one which would have the inverse instance set.

Fixes #18689

[Sean Griffin & anthonynavarre]

There are many ways that things end up getting passed to `concat`. Not
all of those entry points called `flatten` on their input. It seems that
just about every method that is meant to take a single record, or that
splats its input, is meant to also take an array. `concat` is the
earliest point that is common to all of the methods which add records to
the association. Partially fixes #18689

It's under private in Active Model as well.

We've now removed all uses of them across the board. All logic lives on
`WhereClause`.

The code assumes that non-single-value methods mean multi value methods.
That is not the case. We need to change the accessor name, and only
assign an array for multi value methods