OpenBLAS controlling thread logic needs to be smarter when thre are overlapping set/restore pairs in different threads #299

Based on joblib/joblib#1871 (comment)

Scenario:

1. Initial limit is L.
2. Thread 1 limits to M, stores the fact that previously limit was L. Thread 1 becomes controlling thread.
3. Thread 2 tries to limit to N, but fails, stores the fact that previously limit was M.
4. Thread 1 restores to L and releases control of thread.
5. Thread 2 restores to N, acquiring control as it does so.

The straightforward solution here is something like "if a thread failed to set because it wasn't the controlling thread, it shouldn't restore either."