This addresses Bug: 21685.

Details are in the issue. The following is a summary.

Before the fix:

• The active thread drops the GVL and wakes another ready thread on each syscall.
• This causes significant context swiching on real world applications (including rails).
• As a result of the high rate of context switching, multicore machines may experience 10x worse performance on some benchmarks compared to single-core machines.

After the fix:

• The active thread only drops the GVL after a syscall has been blocked for 100usec. This is achieved by adding an extra monitor thread (deferred_wait).
• We see significantly less context switching on benchmarks (1000x less).
• On those same benchmarks, single-core performance improves up to 2x and multi-core performance shows parity with single-core.

Following brief discussion on the issue I am creating this PR to further the discussion.