Simplify dispatcher callbacks to eliminate unnecessary stale thread purging. [Nick Sieger, Pratik Naik]
Signed-off-by: NPratik Naik <pratiknaik@gmail.com>

Signed-off-by: NPratik Naik <pratiknaik@gmail.com>

Signed-off-by: NMichael Koziarski <michael@koziarski.com>

[#1169 state:committed]

After releasing monitor some connection(s) may appear in pool before monitor is re-aquired.
When this happens we'll wait for connection which is already available.
Signed-off-by: NMichael Koziarski <michael@koziarski.com>

Signed-off-by: NJeremy Kemper <jeremy@bitsweat.net>

Signed-off-by: NMichael Koziarski <michael@koziarski.com>

Remove some synchronization that's probably overkill, assuming one doesn't establish connections frequently

- Add connection checkin and checkout callbacks to adapter to allow
  adapter-specific customization of behavior (e.g., JRuby w/ JNDI)

- also clean up some cruft remaining from per-thread connection cache

Remove CachedConnectionPerThread per-thread pooling mechanism in favor of a fixed pool with default maximum of 5 connections

- delegating methods left behind

- Also carry allow_concurrency value through to connection adapter
  (for postgresql)

- use new active support Module#synchronize
- allow_concurrency now switches between a null monitor and a
  regular monitor (defaulting to null monitor to avoid overhead)

So far so good, tests still run clean. Next steps: synchronize connection pool access
and modification, and change allow_concurrency to simply switch a real lock for a
null lock.