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

We're still using it in `where_unscoping`, which will require moving
additional logic.

This will make it easy to add `having_clause` and `join_clause` later.

Yes, I know, I called it a factory so I'm basically the worst person
ever who loves Java and worships the Gang of Four.

Bind values are no longer a thing, so this is unnecessary.

This object being a black box, it knows the details of how to merge
itself with another where clause. This removes all references to where
values or bind values in `Relation::Merger`

The way that bind values are currently stored on Relation is a mess.
They can come from `having`, `where`, or `join`. I'm almost certain that
`having` is actually broken, and calling `where` followed by `having`
followed by `where` will completely scramble the binds.

Joins don't actually add the bind parameters to the relation itself, but
instead add it onto an accessor on the arel AST which is undocumented,
and unused in Arel itself. This means that the bind values must always
be accessed as `relation.arel.bind_values + relation.bind_values`.
Anything that doesn't is likely broken (and tons of bugs have come up
for exactly that reason)

The result is that everything dealing with `Relation` instances has to
know far too much about the internals. The binds are split, combined,
and re-stored in non-obvious ways that makes it difficult to change
anything about the internal representation of `bind_values`, and is
extremely prone to bugs.

So the goal is to move a lot of logic off of `Relation`, and into
separate objects. This is not the same as what is currently done with
`JoinDependency`, as `Relation` knows far too much about its internals,
and vice versa. Instead these objects need to be black boxes that can
have their implementations swapped easily.

The end result will be two classes, `WhereClause` and `JoinClause`
(`having` will just re-use `WhereClause`), and there will be a single
method to access the bind values of a `Relation` which will be
implemented as

```
join_clause.binds + where_clause.binds + having_clause.binds
```

This is the first step towards that refactoring, with the internal
representation of where changed, and an intermediate representation of
`where_values` and `bind_values` to let the refactoring take small
steps. These will be removed shortly.

See 4d7a6229 for the reasoning

The structure of `values[:where]` is going to change, with an
intermediate definition of `where_values` to aid the refactoring.
Accessing `values[:where]` directly messes with that, signficantly.

The array wrapping is no longer necessary, since `where_values` will
always return an array.

[ci skip]

This will allow all types which require no additional handling to use
prepared statements. Specifically, this will allow for `true`, `false`,
`Date`, `Time`, and any custom PG type to use prepared statements. This
also revealed another source of nil columns in bind params, and an
inconsistency in their use.

The specific inconsistency comes from a nested query coming from a
through association, where one of the inversed associations is not
bi-directional.

The stop-gap is to simply construct the column at the site it is being
used. This should simply go away on its own once we use `Attribute` to
represent them instead, since we already have all of the information we
need.

This is to help facilitate future refactorings, as the internal
representation is changed. I'm planning on having `where_values` return
an array that's computed on call, which means that mutation will have no
affect. This is the only remaining place that was mutating (tested by
replacing the method with calling `dup`)

Specifically, the issue is relying on `where_unscoping` mutating the
where values. It does not, however, mutate the bind values, which could
cause an error under certain circumstances. This was not exposed by the
tests, since the only place which would have been affected is unscoping
a boolean, which doesn't go through prepared statements. I had a hard
time getting better test coverage to demonstrate the issue.

This in turn, caused `merge` to go through proper logic, and try to
clear out the binds associated with the unscoped relation, which then
exposed a source of `nil` for the columns, as binds weren't expanding
`{ "posts.id" => 1 }` to `{ "posts" => { "id" => 1 } }`. This has been
fixed.

The bulk of `create_binds` needed to be moved to a separate method,
since the dot notation should not be expanded recursively.

I'm pretty sure this removes a subtle quirk that a ton of code in
`Relation::Merger` is working around, and I suspect that code can be
greatly simplified. However, unraveling that rats nest is no small task.

Minor style changes across the board. Changed an alias to an explicit
method declaration, since the alias will not be documented otherwise.

Allows to use it for any object as an includable module.

Given that this was originally added to normalize an error that would
have otherwise come from the database (inconsistently), it's more
natural for us to raise in `type_cast_for_database`, rather than
`type_cast_from_user`. This way, things like numericality validators can
handle it instead if the user chooses to do so. It also fixes an issue
where assigning an out of range value would make it impossible to assign
a new value later.

This fixes several vague issues, none of which were ever directly
reported, so I have no issue number to give. Places it was mentioned
which I can remember:

- https://github.com/thoughtbot/shoulda-matchers/blob/9ba21381d7caf045053a81f32df7de2f49687820/lib/shoulda/matchers/active_model/allow_value_matcher.rb#L261-L263
- https://github.com/rails/rails/issues/18653#issuecomment-71197026

Fixes #18580.

Fixes #18632

Keeping with our behavior elsewhere in the system, invalid input is
assumed to be `nil`.

Fixes #18629.

This method can be used to see all of the fields on a model which have
been read. This can be useful during development mode to quickly find
out which fields need to be selected. For performance critical pages, if
you are not using all of the fields of a database, an easy performance
win is only selecting the fields which you need. By calling this method
at the end of a controller action, it's easy to determine which fields
need to be selected.

While writing this, I also noticed a place for an easy performance win
internally which I had been wanting to introduce. You cannot mutate a
field which you have not read. Therefore, we can skip the calculation of
in place changes if we have never read from the field. This can
significantly speed up methods like `#changed?` if any of the fields
have an expensive mutable type (like `serialize`)

```
Calculating -------------------------------------
 #changed? with serialized column (before)
                       391.000  i/100ms
 #changed? with serialized column (after)
                         1.514k i/100ms
-------------------------------------------------
 #changed? with serialized column (before)
                          4.243k (± 3.7%) i/s -     21.505k
 #changed? with serialized column (after)
                         16.789k (± 3.2%) i/s -     84.784k
```

With the old implementation, the bind values were created, and then we
search the attributes for `Relation` objects, and merge them. This
completely ignores the order that the actual `where` clause will use. If
all non-relation where parameters are before the relations, it will
work. However, if we query on both a relation and a value, with the
value coming second, it breaks. The order of the hash should not affect
the final query (especially since hashes being ordered is an
implementation detail)

I'm looking to introduce a `WhereClause` class to handle most of this
logic, and this method will eventually move over to there. However, this
intermediate refactoring should make that easier to do.