Adds the ability to save custom types, which type cast to non-primitive
ruby objects.

This mirrors the layout of abstract adapter and puts the definitions
inside the `PostgreSQL` namespace (no longer under the adapter namespace).

/cc @kares

It appears to have been used at some point in the past, but is no longer
used in any meaningful way. Whether a column is considered primary is
a property of the model, not the schema/column. This also removes the
need for yet another layer of caching of the model's schema, and we can
leave that to the schema cache.

The use of default procs was unnessecary, made the code confusing to
follow, and made marshalling needlessly complex.

Columns and injected types no longer have any conditionals based on the
format of SQL type strings! Hooray!

Ideally types will be usable without having to specify a sql type
string, so we should keep the information related to parsing them on the
adapter or another object.

We're going to want all of the benefits of the type map object for
registrations, including block registration and real aliasing. Moves
type name registrations to the adapter, and aliases the OIDs to the
named types

Determining things like precision and scale in postgresql will require
the given blocks to take additional arguments besides the OID.

- Adds the ability to handle additional arguments to `TypeMap`
- Passes the column type to blocks when looking up PG types

The only type that has a scale is decimal. There's a special case where
decimal columns with 0 scale are type cast to integers if the scale is
not specified. Appears to only affect schema dumping.

As we promote these classes to first class concepts, these classes are
starting to gain enough behavior to warrant being moved into their own
files. Many of them will become quite large as we move additional
behavior to the type objects.

  - `extract_precision`, `extract_limit`, and `extract_default` probably need to follow.
  - would be good to remove the delegation `Column#extract_scale`.

/cc @sgrif

All subclasses of column were now delegating `type_cast` to their
injected type object. We can remove the overriding methods, and
generalize it on the `Column` class itself. This also enabled us to
remove several column classes completely, as they no longer had any
meaningful behavior of their own.

Using general types where possible. Several more can go away once
infinity gets figured out.

The `:timestamp` type for columns is unused. All database adapters treat
them as the same database type. All code in `ActiveRecord` which changes
its behavior based on the column's type acts the same in both cases.
However, when the type is passed to code that checks for the `:datetime`
type, but not `:timestamp` (such as XML serialization), the result is
unexpected behavior.

Existing schema definitions will continue to work, and the `timestamp`
type is transparently aliased to `datetime`.

The decision to wrap type registrations in a proc was made for two
reasons.

1. Some cases need to make an additional decision based on the type
  (e.g. a `Decimal` with a 0 scale)
2. Aliased types are automatically updated if they type they point to is
  updated later. If a user or another adapter decides to change the
  object used for `decimal` columns, `numeric`, and `number` will
  automatically point to the new type, without having to track what
  types are aliased explicitly.

Everything else here should be pretty straightforward. PostgreSQL ranges
had to change slightly, since the `simplified_type` method is gone.

Partial revert of c0bfc3f4

Part of #15134. In order to perform typecasting polymorphically, we need
to add another argument to the constructor. The order was chosen to
match the `oid_type` on `PostgreSQLColumn`.

The current behavior is that they are treated as `datetime` normally,
but if they are part of an array, they are treated as `timestamp`. The
only place that seems to be impacted by this is schema dumping, which
shouldn't matter since `t.datetime` and `t.timestamp` are equivalent in
the `PostgreSQL` adapter, anyway.