We don't need separate @class_to_pool and @connection_pool hashes.

* Loop rather than recurse in retrieve_connection_pool
* Key the hash by class rather than class name. This avoids creating
  unnecessary strings.

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

The core of this fix is a threadsafe, fair Queue class.  It is
very similar to Queue in stdlib except that it supports waiting
with a timeout.

The issue this solves is that if several threads are contending for
database connections, an unfair queue makes is possible that a thread
will timeout even while other threads successfully acquire and release
connections.  A fair queue means the thread that has been waiting the
longest will get the next available connection.

This includes a few test fixes to avoid test ordering issues that
cropped up during development of this patch.

An AR ConnectionSpec `wait_timeout` is pre-patch used for three
different things:

* mysql2 uses it for MySQL's own wait_timeout (how long MySQL
  should allow an idle connection before closing it), and
  defaults to 2592000 seconds.
* ConnectionPool uses it for "number of seconds to block and
  wait for a connection before giving up and raising a timeout error",
  default 5 seconds.
* ConnectionPool uses it for the Reaper, for deciding if a 'dead'
  connection can be reaped. Default 5 seconds.

Previously, if you want to change these from defaults, you need
to change them all together. This is problematic _especially_
for the mysql2/ConnectionPool conflict, you will generally _not_
want them to be the same, as evidenced by their wildly different
defaults. This has caused real problems for people #6441 #2894

But as long as we're changing this, forcing renaming the
ConnectionPool key to be more specific, it made sense
to seperate the two ConnectionPool uses too -- these two
types of ConnectionPool timeouts ought to be able to be
changed independently, you won't neccesarily want them
to be the same, even though the defaults are (currently)
the same.

This reverts commit d2901f0f, reversing
changes made to 525839fd.

Conflicts:
	activerecord/test/cases/connection_pool_test.rb

Reason: This change broke the build (http://travis-ci.org/#!/rails/rails/builds/1391490)
and we don't have any solution until now. I asked the author to try to
fix it and open a new pull request.

deprecated clear_stale_active_connections! can call #reap instead of no-op'ing, #reap does the same thing

make active_connection? return true only if there is an open connection in use for the current thread. fixes #5330

The method verify_active_connections! was used in
the old days (up to 2.1 I think) by the dispatcher
to verify the connections, but nowadays we do that
in a different way and this method is obsolete.

Connection pools are 1:1 with pids.

This is the 'top level' connection, inherited by any models that include
ActiveRecord::Model or inherit from ActiveRecord::Base.

it clear which methods are synchronized.