R1:62b06f0f5dd7

Fix a memory leak in PhabricatorGlobalLock

Summary:
We currently cache all connections in LiskDAO so we can roll back transactions when fixtured unit tests complete.

Since we establish a new connection wrapper each time we establish a global lock, this cache currently grows without bound.

Instead, pool global lock connections so we never have more than the largest number of locks we've held open at once (in PullLocalDaemon, always 1).

Another way to solve this is probably to add an "onclose" callback…