Override callback doesn't work anymore.

This reverts commit 3420a145, reversing
changes made to afb66a5a.

We pretty frequently get bug reports that "dirty is broken inside of
after callbacks". Intuitively they are correct. You'd expect
`Model.after_save { puts changed? }; model.save` to do the same thing as
`model.save; puts model.changed?`, but it does not.

However, changing this goes much farther than just making the behavior
more intuitive. There are a _ton_ of places inside of AR that can be
drastically simplified with this change. Specifically, autosave
associations, timestamps, touch, counter cache, and just about anything
else in AR that works with callbacks have code to try to avoid "double
save" bugs which we will be able to flat out remove with this change.

We introduce two new sets of methods, both with names that are meant to
be more explicit than dirty. The first set maintains the old behavior,
and their names are meant to center that they are about changes that
occurred during the save that just happened. They are equivalent to
`previous_changes` when called outside of after callbacks, or once the
deprecation cycle moves.

The second set is the new behavior. Their names imply that they are
talking about changes from the database representation. The fact that
this is what we really care about became clear when looking at
`BelongsTo.touch_record` when tests were failing. I'm unsure that this
set of methods should be in the public API. Outside of after callbacks,
they are equivalent to the existing methods on dirty.

Dirty itself is not deprecated, nor are the methods inside of it. They
will only emit the warning when called inside of after callbacks. The
scope of this breakage is pretty large, but the migration path is
simple. Given how much this can improve our codebase, and considering
that it makes our API more intuitive, I think it's worth doing.

Change to use a more realistic example and not giving the impression that destroy_all is preferred way to destroy related records.
This example just wants to demonstrate callback behaviour.

[ci skip]

After fb898e98, the `before_destroy` had some code that used
SQL interpolation left over. Don't think we should promote
that even if the values aren't directly from user input.

Passing conditions to `#destroy_all` was deprecated in c82c5f8f.

In the doc the `dependent` option was set with: `dependent: destroy`.
This is not working because destroy would call the method of the activerecord::base object.

The right way is: `dependent: :destroy`

The focus of this change is to make the API more accessible.
References to method and classes should be linked to make it easy to
navigate around.

This patch makes exzessiv use of `rdoc-ref:` to provide more readable
docs. This makes it possible to document `ActiveRecord::Base#save` even
though the method is within a separate module
`ActiveRecord::Persistence`. The goal here is to bring the API closer to
the actual code that you would write.

This commit only deals with Active Record. The other gems will be
updated accordingly but in different commits. The pass through Active
Record is not completely finished yet. A follow up commit will change
the spots I haven't yet had the time to update.

/cc @fxn

`+` doesn't work around content with spaces fallback `<tt>`.

Deep down in the association internals, we're calling `destroy!` rather
than `destroy` when handling things like `dependent` or autosave
association callbacks. Unfortunately, due to the structure of the code
(e.g. it uses callbacks for everything), it's nearly impossible to pass
whether to call `destroy` or `destroy!` down to where we actually need
it.

As such, we have to do some legwork to handle this. Since the callbacks
are what actually raise the exception, we need to rescue it in
`ActiveRecord::Callbacks`, rather than `ActiveRecord::Persistence` where
it matters. (As an aside, if this code wasn't so callback heavy, it
would handling this would likely be as simple as changing `destroy` to
call `destroy!` instead of the other way around).

Since we don't want to lose the exception when `destroy!` is called (in
particular, we don't want the value of the `record` field to change to
the parent class), we have to do some additional legwork to hold onto it
where we can use it.

Again, all of this is ugly and there is definitely a better way to do
this. However, barring a much more significant re-architecting for what
I consider to be a reletively minor improvement, I'm willing to take
this small hit to the flow of this code (begrudgingly).

This reverts commit bdc1d329.

Before:
Calculating -------------------------------------
                        22.000  i/100ms
-------------------------------------------------
                        229.700  (± 0.4%) i/s -      1.166k
Total Allocated Object: 9939

After:
Calculating -------------------------------------
                        24.000  i/100ms
-------------------------------------------------
                        246.443  (± 0.8%) i/s -      1.248k
Total Allocated Object: 7939

```
begin
  require 'bundler/inline'
rescue LoadError => e
  $stderr.puts 'Bundler version 1.10 or later is required. Please update your Bundler'
  raise e
end

gemfile(true) do
  source 'https://rubygems.org'
  # gem 'rails', github: 'rails/rails', ref: 'bdc1d329'
  gem 'rails', github: 'rails/rails', ref: 'd2876141'
  gem 'arel', github: 'rails/arel'
  gem 'sqlite3'
  gem 'benchmark-ips'
end

require 'active_record'
require 'benchmark/ips'

ActiveRecord::Base.establish_connection('sqlite3::memory:')

ActiveRecord::Migration.verbose = false

ActiveRecord::Schema.define do
  create_table :users, force: true do |t|
    t.string :name, :email
    t.boolean :admin
    t.timestamps null: false
  end
end

class User < ActiveRecord::Base
  default_scope { where(admin: true) }
end

admin = true

1000.times do
  attributes = {
    name: "Lorem ipsum dolor sit amet, consectetur adipiscing elit.",
    email: "foobar@email.com",
    admin: admin
  }

  User.create!(attributes)

  admin = !admin
end

GC.disable

Benchmark.ips(5, 3) do |x|
  x.report { User.all.to_a }
end

key =
  if RUBY_VERSION < '2.2'
    :total_allocated_object
  else
    :total_allocated_objects
  end

before = GC.stat[key]
User.all.to_a
after = GC.stat[key]
puts "Total Allocated Object: #{after - before}"
```

This reverts commit 796cab45.

Prior to this commit if you defined a bidirectional relationship
between two models with destroy dependencies on both sides, a call to
`destroy` would result in an infinite callback loop.

Take the following relationship.

    class Content < ActiveRecord::Base
      has_one :content_position, dependent: :destroy
    end

    class ContentPosition < ActiveRecord::Base
      belongs_to :content, dependent: :destroy
    end

Calling `Content#destroy` or `ContentPosition#destroy` would result in
an infinite callback loop.

This commit changes the behaviour of `ActiveRecord::Callbacks#destroy`
so that it guards against subsequent callbacks.

Thanks to @zetter for demonstrating the issue with failing tests[1].

[1] rails#13609

Before this commit, returning `false` in an ActiveRecord `before_` callback
such as `before_create` would halt the callback chain.

After this commit, the behavior is deprecated: will still work until
the next release of Rails but will also display a deprecation warning.

The preferred way to halt a callback chain is to explicitly `throw(:abort)`.

This will avoid naming clash with user defined methods

Inspired by @tenderlove's work in
c363fff2, this reduces the number of
strings allocated when running callbacks for ActiveRecord instances. I
measured that using this script:

```
require 'objspace'
require 'active_record'
require 'allocation_tracer'

ActiveRecord::Base.establish_connection adapter: "sqlite3",
                                        database: ":memory:"

ActiveRecord::Base.connection.instance_eval do
  create_table(:articles) { |t| t.string :name }
end

class Article < ActiveRecord::Base; end
a = Article.create name: "foo"
a = Article.find a.id

N = 10
result = ObjectSpace::AllocationTracer.trace do
  N.times { Article.find a.id }
end

result.sort.each do |k,v|
  p k => v
end
puts "total: #{result.values.map(&:first).inject(:+)}"
```

When I run this against master and this branch I get this output:

```
pete@balloon:~/projects/rails/activerecord$ git checkout master
M Gemfile
Switched to branch 'master'
pete@balloon:~/projects/rails/activerecord$ bundle exec ruby benchmark_allocation_with_callback_send.rb > allocations_before
pete@balloon:~/projects/rails/activerecord$ git checkout remove-dynamic-send-on-built-in-callbacks
M Gemfile
Switched to branch 'remove-dynamic-send-on-built-in-callbacks'
pete@balloon:~/projects/rails/activerecord$ bundle exec ruby benchmark_allocation_with_callback_send.rb > allocations_after
pete@balloon:~/projects/rails/activerecord$ diff allocations_before allocations_after
39d38
<
{["/home/pete/projects/rails/activesupport/lib/active_support/callbacks.rb",
81]=>[40, 0, 0, 0, 0, 0]}
42c41
< total: 630
---
> total: 590

```

In addition to this, there are two micro-optimizations present:

* Using `block.call if block` vs `yield if block_given?` when the block was being captured already.

```
pete@balloon:~/projects$ cat benchmark_block_call_vs_yield.rb
require 'benchmark/ips'

def block_capture_with_yield &block
  yield if block_given?
end

def block_capture_with_call &block
  block.call if block
end

def no_block_capture
  yield if block_given?
end

Benchmark.ips do |b|
  b.report("block_capture_with_yield") { block_capture_with_yield }
  b.report("block_capture_with_call") { block_capture_with_call }
  b.report("no_block_capture") { no_block_capture }
end
pete@balloon:~/projects$ ruby benchmark_block_call_vs_yield.rb
Calculating -------------------------------------
block_capture_with_yield
                        124979 i/100ms
block_capture_with_call
                        138340 i/100ms
    no_block_capture    136827 i/100ms
-------------------------------------------------
block_capture_with_yield
                      5703108.9 (±2.4%) i/s -   28495212 in   4.999368s
block_capture_with_call
                      6840730.5 (±3.6%) i/s -   34169980 in   5.002649s
    no_block_capture  5821141.4 (±2.8%) i/s -   29144151 in   5.010580s
```

* Defining and calling methods instead of using send.

```
pete@balloon:~/projects$ cat benchmark_method_call_vs_send.rb
require 'benchmark/ips'

class Foo
  def tacos
    nil
  end
end

my_foo = Foo.new

Benchmark.ips do |b|
  b.report('send') { my_foo.send('tacos') }
  b.report('call') { my_foo.tacos }
end
pete@balloon:~/projects$ ruby benchmark_method_call_vs_send.rb
Calculating -------------------------------------
                send     97736 i/100ms
                call    151142 i/100ms
-------------------------------------------------
                send  2683730.3 (±2.8%) i/s -   13487568 in   5.029763s
                call  8005963.9 (±2.7%) i/s -   40052630 in   5.006604s
```

The result of this is making typical ActiveRecord operations slightly faster:

https://gist.github.com/phiggins/e46e51dcc7edb45b5f98

 This is to ensure that they are not accidentally called by the app code.
 They are renamed to _create_record and _update_record respectively.
 Closes #11645

Callback caller class uses `after_initialize`,
but Callback callee defines `after_find`.

Current sample code causes following error.

  NoMethodError: undefined method `after_initialize' for #<EncryptionWrapper:0x007fe4931fa5c0>

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