variables [ci skip]
Thanks to @matthewd for the excellent feedback! ❤

This partially reverts commit 1a203d5e.

Reason: It was committed by accident

There's a lot more that can be moved to these, but this felt like a good
place to introduce the object. Plans are:

- Remove all knowledge of type casting from the columns, beyond a
  reference to the cast_type
- Move type_cast_for_database to these objects
- Potentially make them mutable, introduce a state machine, and have
  dirty checking handled here as well
- Move `attribute`, `decorate_attribute`, and anything else that
  modifies types to mess with this object, not the columns hash
- Introduce a collection object to manage these, reduce allocations, and
  not require serializing the types

We guarantee that `model.value` does not change after
`model.save && model.reload`. This requires type casting user input for
non-string types.

We'd spend a lot of time calling `hash` and `eql?` on the join model,
which has no primary key. Calling `id` with no primary key is a really
slow way to get back `nil`, so we can improve the performance there.
However, even with the escape clause, we *still* weren't getting high
enough performance, as we were checking the primary key too much. `hash`
will always return `nil.hash` for records with no id, and `==` will
always return `false`. We can optimize those cases in the HABTM join
model.

Additionally:
* It changes `purge` task on `sqlite3` adapter to recreate database file, to
be consistent with other adapters.
* Adds `purge` step when loading from `schema.rb`

The column name given by the adapter doesn't include the table
namespace, so going through the hashed version of the result set causes
overridden keys.

Fixes #15649

We're not longer using `ipaddr` in schema dumper

The original patch that added this concept can be found
[here](https://web.archive.org/web/20090601022739/http://dev.rubyonrails.org/ticket/9767).
The current default behavior is to cache everything except serialized
columns, unless the user specified otherwise. If anyone were to specify
otherwise, many types would actually be completely broken. Still, the
method is left in place with a deprecation warning in case anyone is
actually still calling this method.

We are no longer including column objects in YAML serialization, thanks
to https://github.com/rails/rails/pull/15621

add_constraints is complicated and difficult to read. This is the
beginning of a long process of refactoring this code. First step:
moved the join keys out of AssociationScope and into reflection.

We then don't need to call `reflection` because now reflection is
`self`.

`foreign_key` must be named something else because reflection already has a
`foreign_key` method and when passed into `JoinKeys` it was getting the
wrong assignment. `reflection_foreign_key` seemed to be an appropriate name.
I also named `key` `reflection_key` to match `reflection_foreign_key`.

We never want result types to override column types, and
`decorate_columns` can only affect column types. No need to go through
the decoration multiple times, we can just exclude the column types from
the result types instead.

On MySQL and PostgreSQL, the adapter does not type cast virtual columns
for us.

Reflection has an available method that is used to check if the
reflection is a collection. Any :has_many macro is considered a
collection and `collection?` should be used instead of
`macro == :has_many`.

By having the `:has_and_belongs_to_many` macro in the `@collection`
we are punishing `:has_many` associations because it has to allocate
the array and check the macro.

@collection is returned to `macro == :has_many` and a new reflection
class `HABTMReflection` is created to handle this case instead.

Instead of checking for `macro == :has_one` throughout the
codebase we can create a `has_one?` method to match the `belongs_to?`,
`polymorphic?` and other methods.

This reverts commit dd3ea171 and add a
regression test.

Fixes #15418

In some cases there is a difference between the two, we should always
be doing one or the other. For convenience, `type_cast` is still a
private method on type, so new types that do not need different behavior
don't need to implement two methods, but it has been moved to private so
it cannot be used accidentally.

We should be able to rely on this object implenting the full type
interface.

Integration tests are inside protected_attributes test suite.

Fixes #15496
Fixes rails/protected_attributes#35

- The following is now true for all types, all the time
  - `model.attribute_before_type_cast == given_value`
  - `model.attribute == model.save_and_reload.attribute`
  - `model.attribute == model.dup.attribute`
  - `model.attribute == YAML.load(YAML.dump(model)).attribute`
- Removes the remaining types implementing `type_cast_for_write`
- Simplifies the implementation of time zone aware attributes
- Brings tz aware attributes closer to being implemented as an attribute
  decorator
- Adds additional point of control for custom types

`[].max # => nil`

The definition of `write_attribute` in dirty checking ultimately leads
to the columns calling `type_cast` on the value to perform the
comparison. However, this is a potentially expensive computation that we
cache when it occurs in `read_attribute`. The only case that we need the
non-type-cast form is for numeric, so we pass that through as well
(something I'm looking to remove in the future).

This also reduces the number of places that manually access various
stages in an attribute's type casting lifecycle, which will aid in one
of the larger refactorings that I'm working on.

Reduces the number of places that care about the internals of how we
store and type cast attributes. We do not need to go through the
dup/freeze dance, as you couldn't have saved a frozen new record anyway,
and that is the only time we would end up modifying the frozen hash.

There is no way to have an instance of an Active Record model where
`has_attribute?(self.class.primary_key)` returns false. The record is
always initialized in such a way that `@raw_attributes` will have an id
key with nil for the value.