Nearly completely implemented in terms of custom properties.
`_before_type_cast` now stores the raw serialized string consistently,
which removes the need to keep track of "state". The following is now
consistently true:

- `model.serialized == model.reload.serialized`
- A model can be dumped and loaded infinitely without changing
- A model can be saved and reloaded infinitely without changing

`@attributes` was actually used for `_before_type_cast` and friends,
while `@attributes_cache` is the type cast version (and caching is the
wrong word there, but I'm working on removing the conditionals around
that). I opted for `@raw_attributes`, because `_before_type_cast` is
also semantically misleading. The values in said hash are in the state
given by the form builder or database, so raw seemed to be a good word.

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`.

see: https://github.com/blog/1406-namespaced-gists

This improves memory and performance without having to use symbols which
present DoS problems. Thanks @headius and @tenderlove for the
suggestion.

This was originally committed in
f1765019, and then reverted in
d3494903 due to it causing problems in a
real application. This second attempt should solve that.

Benchmark
---------

require 'active_record'
require 'benchmark/ips'

ActiveRecord::Base.establish_connection(adapter: 'sqlite3', database: ':memory:')

class Post < ActiveRecord::Base
  connection.create_table :posts, force: true do |t|
    t.string :name
  end
end

post = Post.create name: 'omg'

Benchmark.ips do |r|
  r.report('Post.new')          { Post.new name: 'omg' }
  r.report('post.name')         { post.name }
  r.report('post.name=')        { post.name = 'omg' }
  r.report('Post.find(1).name') { Post.find(1).name }
end

Before
------

Calculating -------------------------------------
            Post.new      1419 i/100ms
           post.name      7538 i/100ms
          post.name=      3024 i/100ms
   Post.find(1).name       243 i/100ms
-------------------------------------------------
            Post.new    20637.6 (±12.7%) i/s -     102168 in   5.039578s
           post.name  1167897.7 (±18.2%) i/s -    5186144 in   4.983077s
          post.name=    64305.6 (±9.6%) i/s -     317520 in   4.998720s
   Post.find(1).name     2678.8 (±10.8%) i/s -      13365 in   5.051265s

After
-----

Calculating -------------------------------------
            Post.new      1431 i/100ms
           post.name      7790 i/100ms
          post.name=      3181 i/100ms
   Post.find(1).name       245 i/100ms
-------------------------------------------------
            Post.new    21308.8 (±12.2%) i/s -     105894 in   5.053879s
           post.name  1534103.8 (±2.1%) i/s -    7634200 in   4.979405s
          post.name=    67441.0 (±7.5%) i/s -     337186 in   5.037871s
   Post.find(1).name     2681.9 (±10.6%) i/s -      13475 in   5.084511s

This reverts commit f1765019.

In the end I think the pain of implementing this seamlessly was not
worth the gain provided.

The intention was that it would allow plain ruby objects that might not
live in your main application to be subclassed and have persistence
mixed in. But I've decided that the benefit of doing that is not worth
the amount of complexity that the implementation introduced.

This reverts commit 83846838.

I think it's going to be too much pain to try to transition the
:active_record load hook from executing against Base to executing
against Model.

For example, after Model is included in Base, and modules included in
Model will no longer get added to the ancestors of Base.

So plugins which wish to be compatible with both Model and Base should
use the :active_record_model load hook which executes *before* Base gets
loaded.

In general, ActiveRecord::Model is an advanced feature at the moment and
probably most people will continue to inherit from ActiveRecord::Base
for the time being.

This improves memory and performance without having to use symbols which
present DoS problems. Thanks @headius and @tenderlove for the
suggestion.

Benchmark
---------

require 'active_record'
require 'benchmark/ips'

ActiveRecord::Base.establish_connection(adapter: 'sqlite3', database:
':memory:')

class Post < ActiveRecord::Base
  connection.create_table :posts, force: true do |t|
    t.string :name
  end
end

post = Post.create name: 'omg'

Benchmark.ips do |r|
  r.report('Post.new')          { Post.new name: 'omg' }
  r.report('post.name')         { post.name }
  r.report('post.name=')        { post.name = 'omg' }
  r.report('Post.find(1).name') { Post.find(1).name }
end

Before
------

Calculating -------------------------------------
            Post.new      1419 i/100ms
           post.name      7538 i/100ms
          post.name=      3024 i/100ms
   Post.find(1).name       243 i/100ms
-------------------------------------------------
            Post.new    20637.6 (±12.7%) i/s -     102168 in   5.039578s
           post.name  1167897.7 (±18.2%) i/s -    5186144 in   4.983077s
          post.name=    64305.6 (±9.6%) i/s -     317520 in   4.998720s
   Post.find(1).name     2678.8 (±10.8%) i/s -      13365 in   5.051265s

After
-----

Calculating -------------------------------------
            Post.new      1431 i/100ms
           post.name      7790 i/100ms
          post.name=      3181 i/100ms
   Post.find(1).name       245 i/100ms
-------------------------------------------------
            Post.new    21308.8 (±12.2%) i/s -     105894 in   5.053879s
           post.name  1534103.8 (±2.1%) i/s -    7634200 in   4.979405s
          post.name=    67441.0 (±7.5%) i/s -     337186 in   5.037871s
   Post.find(1).name     2681.9 (±10.6%) i/s -      13475 in   5.084511s

This reverts commit 86c3dfbd.

Conflicts:
	activerecord/lib/active_record/attribute_methods/read.rb

Reason: whilst this increased performance, it also presents a DoS risk
via memory exhaustion if users were allowing user input to dictate the
arguments of read/write_attribute. I will investigate alternative ways
to cut down on string allocations here.

This is purely a performance optimisation.

See https://gist.github.com/3552829

This is a performance/GC optimisation.

In theory, this could be optimised by the implementation (last time I
checked, this would have no effect on JRuby). But in practise, this make
attribute access faster.

Get rid of ActiveModel::Configuration, make better use of
ActiveSupport::Concern + class_attribute, etc.