Session tracing has too high an impact on pillowfight 80/20 throughput

Spun out of MB-28409, analysis there has shown that completely disabling session tracing increases throughput by ~8%. It only has a 1% budget by design, so it currently exceeds that significantly:

I re-ran build 1733 on the R80/R20 pillowfight workload (http://perf.jenkins.couchbase.com/job/triton/12744/), and also a toy build based on 1733 but with tracing disabled (http://review.couchbase.org/#/c/91751/):

Build	Job	Throughput	Delta
1733	triton 12744	3183680	-
11733 (no tracing)	triton 12747	3432963	+7.8%

This completely disabled tracing, including recording timings for slow ops (i.e. work which is done even when the client doesn't opt-in to tracing).

We need to reduce this down to below 1%.