The way that bind values are currently stored on Relation is a mess.
They can come from `having`, `where`, or `join`. I'm almost certain that
`having` is actually broken, and calling `where` followed by `having`
followed by `where` will completely scramble the binds.

Joins don't actually add the bind parameters to the relation itself, but
instead add it onto an accessor on the arel AST which is undocumented,
and unused in Arel itself. This means that the bind values must always
be accessed as `relation.arel.bind_values + relation.bind_values`.
Anything that doesn't is likely broken (and tons of bugs have come up
for exactly that reason)

The result is that everything dealing with `Relation` instances has to
know far too much about the internals. The binds are split, combined,
and re-stored in non-obvious ways that makes it difficult to change
anything about the internal representation of `bind_values`, and is
extremely prone to bugs.

So the goal is to move a lot of logic off of `Relation`, and into
separate objects. This is not the same as what is currently done with
`JoinDependency`, as `Relation` knows far too much about its internals,
and vice versa. Instead these objects need to be black boxes that can
have their implementations swapped easily.

The end result will be two classes, `WhereClause` and `JoinClause`
(`having` will just re-use `WhereClause`), and there will be a single
method to access the bind values of a `Relation` which will be
implemented as

```
join_clause.binds + where_clause.binds + having_clause.binds
```

This is the first step towards that refactoring, with the internal
representation of where changed, and an intermediate representation of
`where_values` and `bind_values` to let the refactoring take small
steps. These will be removed shortly.

This is no longer required now that we are injecting a type caster
object into the Arel table, with the exception of uniqueness
validations. Since it calls `ConnectionAdapter#type_cast`, the value has
already been cast for the database. We don't want Arel to attempt to
cast it further, so we need to continue wrapping it in a quoted node.
This can potentially go away when this validator is refactored to make
better use of `where` or the predicate builder.

Part of the larger refactoring to remove type casting from Arel. We can
inform it that we already have the right type by wrapping the value in
an `Arel::Nodes::Quoted`. This commit can be reverted when we have
removed type casting from Arel in Rail 5.1

Construction of relations can be a hotspot, we don't want to create one
of these in the constructor. This also allows us to do more expensive
things in the predicate builder's constructor, since it's created once
per AR::Base subclass

We have several test cases on "tricky" types that are essentially
testing that `update_all` goes through the same type casting behavior as
a normal assignment + save. We recently had another case to add this
test for another type in https://github.com/rails/rails/pull/12742.

Rather than testing this separately for every type which is "tricky"
when round tripping, let's instead have a fairly exhaustive test that
ensures we're getting the correct values at every step for `update_all`.
Given the structure of the code now, we can be confident that if the
type is correct, and `update_all` is type casting correctly, we're going
to get the right behavior for all types.

Outer joins were being built on the root relation klass rather than the
one specified in the join dependency root

The original code ignores the `false` value because `false.blank? # => true`.

Stop special-casing Arel::Nodes as exempt from reference scanning in
order. Instead, only scan order values that are strings for a table
reference.

fixes #10669

While joining_values special treatment is given to string values.
By flattening the array it ensures that string values are detected
as strings and not arrays.

Object#respond_to? returns singletons and thus we inherit that contract.
The implementation of the predicate is good, but the test is only
checking boolean semantics, which in this case is not enough.

fixes #4208

If a query selects only a few columns and gives custom names to
those columns then respond_to? was returning true for the non
selected columns. However calling those non selected columns
raises exception.

    post = Post.select("'title' as post_title").first

In the above case when `post.body` is invoked then an exception is
raised since `body` attribute is not selected. Howevere `respond_to?`
did not behave correctly.

    pos.respond_to?(:body) #=> true

Reason was that Active Record calls `super` to pass the call to
Active Model and all the columns are defined on Active Model.

Fix is to actually check if the data returned from the db contains
the data for column in question.

Fixes #3002. Also see #5494.

```
class Comment < ActiveRecord::Base
  belongs_to :post
end

class Author < ActiveRecord::Base
  has_many :posts
end

class Post < ActiveRecord::Base
  belongs_to :author
  has_many :comments
end
```

`Comment.joins(:post).merge(Post.joins(:author).merge(Author.where(:name => "Joe Blogs"))).all` would
fail with `ActiveRecord::ConfigurationError: Association named 'author' was not found on Comment`.

It is failing because `all` is being called on relation which looks like this after all the merging:
`{:joins=>[:post, :author], :where=>[#<Arel::Nodes::Equality: ....}`. In this relation all the context that
`Post` was joined with `Author` is lost and hence the error that `author` was not found on `Comment`.

Ths solution is to build JoinAssociation when two relations with join information are being merged. And later
while building the arel use the  previously built `JoinAssociation` record in `JoinDependency#graft` to
build the right from clause.

Thanks to Jared Armstrong (https://github.com/armstrjare) for most of the work. I ported it to make it
compatible with new code base.

The similarity of `Relation#uniq` to `Array#uniq` is confusing. Since our
Relation API is close to SQL terms I renamed `#uniq` to `#distinct`.

There is no deprecation. `#uniq` and `#uniq!` are aliases and will continue
to work. I also updated the documentation to promote the use of `#distinct`.

Fixes #9275.

When `#order` is called with a Symbol this patch will prepend the quoted_table_name.
Before the postgresql adapter failed to build queries containg a join and an order
with a symbol.

This expansion happens for all adapters.

At present, ActiveRecord::Delegation compiles delegation methods on a
global basis. The compiled methods apply to all subsequent Relation
instances. This creates several problems:

1) After Post.all.recent has been called, User.all.respond_to?(:recent)
   will be true, even if User.all.recent will actually raise an error due
   to no User.recent method existing. (See #8080.)

2) Depending on the AR class, the delegation should do different things.
   For example, if a Post.zip method exists, then Post.all.zip should call
   it. But this will then result in User.zip being called by a subsequent
   User.all.zip, even if User.zip does not exist, when in fact
   User.all.zip should call User.all.to_a.zip. (There are various
   variants of this problem.)

We are creating these compiled delegations in order to avoid method
missing and to avoid repeating logic on each invocation.

One way of handling these issues is to add additional checks in various
places to ensure we're doing the "right thing". However, this makes the
compiled methods signficantly slower. In which case, there's almost no
point in avoiding method_missing at all. (See #8127 for a proposed
solution which takes this approach.)

This is an alternative approach which involves creating a subclass of
ActiveRecord::Relation for each AR class represented. So, with this
patch, Post.all.class != User.all.class. This means that the delegations
are compiled for and only apply to a single AR class. A compiled method
for Post.all will not be invoked from User.all.

This solves the above issues without incurring significant performance
penalties. It's designed to be relatively seamless, however the downside
is a bit of complexity and potentially confusion for a user who thinks
that Post.all and User.all should be instances of the same class.

Benchmark
---------

require 'active_record'
require 'benchmark/ips'

class Post < ActiveRecord::Base
  establish_connection adapter: 'sqlite3', database: ':memory:'
  connection.create_table :posts

  def self.omg
    :omg
  end
end

relation = Post.all

Benchmark.ips do |r|
  r.report('delegation')   { relation.omg }
  r.report('constructing') { Post.all }
end

Before
------

Calculating -------------------------------------
          delegation      4392 i/100ms
        constructing      4780 i/100ms
-------------------------------------------------
          delegation   144235.9 (±27.7%) i/s -     663192 in   5.038075s
        constructing   182015.5 (±21.2%) i/s -     850840 in   5.005364s

After
-----

Calculating -------------------------------------
          delegation      6677 i/100ms
        constructing      6260 i/100ms
-------------------------------------------------
          delegation   166828.2 (±34.2%) i/s -     754501 in   5.001430s
        constructing   116575.5 (±18.6%) i/s -     563400 in   5.036690s

Comments
--------

Bear in mind that the standard deviations in the above are huge, so we
can't compare the numbers too directly. However, we can conclude that
Relation construction has become a little slower (as we'd expect), but
not by a huge huge amount, and we can still construct a large number of
Relations quite quickly.

This was requested by DHH to allow creating of one's own custom
association macros.

For example:

    module Commentable
      def has_many_comments(extra)
        has_many :comments, -> { where(:foo).merge(extra) }
      end
    end

    class Post < ActiveRecord::Base
      extend Commentable
      has_many_comments -> { where(:bar) }
    end

Record.from("(#{sub_query.to_sql})")    -> Record.from(sub_query)
Record.from("(#{sub_query.to_sql}) a")  -> Record.from(sub_query, :a)