prange() scheduling is suboptimal when workload varies between iterations

[seibert]

When using prange() to parallelize a loop with a body that takes highly variable amounts of time per iteration, the scheduler will tend to leave some threads idle. This is because the iteration range is divided between the threads upfront to maximize chunk size. If the chunk size were smaller and fed into a work queue, it would allow better load balancing between the worker threads.

This is also likely something which becomes more efficient when we have TBB to do the thread dispatch.

See email thread: https://groups.google.com/a/continuum.io/forum/#!msg/numba-users/ZDWRtE_U2Y8/S-SbZt6YBQAJ

[anton-malakhov]

Relates to #2245