Allows seeing directly how long was spent:
- running up to putting bytes on the wire
- between sending and receiving
- between receiving and finishing

- More robustly run processes in the background
- Detect and use pypy if available
- Add a --latency_profile flag to control whether to run the latency profiling stuff

Current latency profiles have their tails dominated by writing latency
logs, which is hugely undesirable.

Now when a thread log fills up, push it to a background thread to write
to disk. At shutdown, wait for all latency traces to be flushed.