Fix a race condition that could cause lockup

Bug: 22676664

If a task is queued on a worker thread that is busy
processing a task AND a trim memory signal comes along
prior to the initial task being completed then the
tasks that were queued prior to the exit request will
never be processed.

This is bad, as when those futures are used later on, such
as in trying to delete them from the cache, they will block
waiting for the orphaned task to be finished.

Simple fix is to just never stop the worker threads. Increase
RAM usage slightly as we won't reclaim that stack space, but
this cost is fairly small

Change-Id: I818b011f9bc6ba06942fdad81c27656bd97b42d4