More efficient concurrent queues

[lindahua]

The current implementation of concurrent_queue relies on a native synchronization mechanism, i.e. lock the entire queue structure whenever it pushes or pops. This leads to limited scalability. There are much more efficient implementation for MPMC Queues (e.g. Dmitry's lock free implementation that relies purely on atomics).

Plan:

• Rename concurrent_queue to synchronized_queue, which reflects the implementation more accurately.
• Implement a more efficient bounded_mpmc_queue (MPMC means Multi-Producer-Multi-Consumer).
• Implement a task queue on top.

References:

• Dmitry Vyukov's blog on MPMC Queue implementation
• Folly MPMCQueue
• rigtorp/MPMCQueue
• A high performance thread pool based on bounded MPMC Queue