The :singleton-method: directive is printed in the docs.
This directive is not necessary as the method is documentable just fine.

Mostly delegation to start, but we can start moving a lot of behavior in
bulk to this object.

Now the following case will work fine

    class Tag < ActiveRecord::Base
    end

    class Publisher::Article < ActiveRecord::Base
      has_and_belongs_to_many :tags
    end

Fixes #15761

`Type::Integer.new.type_cast('') # => nil`, we do not need a special
case to handle this, `nil => ''` already returns false. The only case we
need to handle is `0 => 'wibble'` should be changed, while `0 => '0'`
should not.

We're never going to be able to use the attribute object here, however,
so let's just accept the ugly demeter violation here for now.

Remove test cases which were either redundant with other tests in the
file, or were actually testing the type objects (which are tested
elsewhere)

Moves towards removing type casting knowledge from the column entirely

Also takes a step towards supporting types which use a character other
than ',' for the delimiter (`box` is the only built in type for which
this is the case)

If we want to have type decorators mess with the attribute, but not the
column, we need to stop type casting on the column. Where possible, we
changed the tests to test the value of `column_defaults`, which is
public API. `Column#default` is not.

The case where we have a column object, but don't have a type cast
method involves type casting the default value when changing the schema.
We get one of the column definition structs instead. That is a case that
I'm trying to remove overall, but in the short term, we can achieve the
same behavior without needing to pass the adapter to the array type by
creating a fake type that proxies to the adapter.

HStore columns come back from the database separated by a comma and a
space, not just a comma. We need to mirror that behavior since we
compare the two values.

Also adds a regression test against JSON to ensure we don't have the
same bug there.

We no longer need to "init changed attributes" from the initializer,
either, as there is no longer a case where a given value would differ
from the default, but would not already be marked as changed.

If the through record gets created in an `after_create` hook that is
defined before the association is defined (therefore after its
`after_create` hook) get saved twice. This ensures that the through
records are created only once, regardless of the order of the hooks.

This refactoring revealed the need for another form of decoration, which
takes a proc to select which it applies to (There's a *lot* of cases
where this form can be used). To avoid duplication, we can re-implement
the old decoration in terms of the proc-based decoration.

The reason we're `instance_exec`ing the matcher is for cases such as
time zone aware attributes, where a decorator is defined in a parent
class, and a method called in the matcher is overridden by a child
class. The matcher will close over the parent, and evaluate in its
context, which is not the behavior we want.

Before, calling `size` would only work if it skipped the cache, and
would return a different result from the cache, but only if:

- The association was previously loaded
- Or you called size previously
- But only if the size was 0 when you called it

This ensures that the counter is appropriately updated in memory.

to "without replacement"

We've stopped using it internally, in favor of polymorphism. So should
you!

Fixes #15705.

This fixes #15496