retrieve_connection_pool

Retrieving the connection pool happens a lot so we cache it in @class_to_pool. This makes retrieving the connection pool O(1) once the process is warm. When a connection is established or removed, we invalidate the cache.

Ideally we would use fetch here, as class_to_pool may sometimes be nil. However, benchmarking (gist.github.com/jonleighton/3552829) showed that fetch is significantly slower than []. So in the nil case, no caching will take place, but that's ok since the nil case is not the common one that we wish to optimise for.