Summary:
There was a problem with DeterministicSchedule when threads issued shared access stores/loads to atomics during thread local destructors. Thread local destructors run after a thread terminates, and therefore after all of the DSched cleanup work in beforeThreadExit() is already complete.

The new DSched::joinAll() functions works around this problem by blocking the parent until all child threads are blocked at the end of beforeThreadExit(). The parent then lets each child thread proceed one at a time through the thread local destructor with a paired sem_post and .join().

Reviewed By: djwatson

Differential Revision: D13381250

fbshipit-source-id: 80ab39b3f2f71ee69dfbaec9950dfc5908874353