Evaluate new metric engine performance

[akoshelev]

I want to see whether using thread-local storages actually improves query performance

[akoshelev]

I had to fix #1345 first because it turned out we've been running with global metric collector all the time, but after turning it off I got some numbers I can share here

[akoshelev]

Overall the impact of using the old metric engine seems to be around 10%. For 1M input, it takes 10 mins with metrics enabled and 9 mins with metrics disabled. The new metric engine does not seem to have any impact on runtime

Input size\Time (s)	No metrics	Metrics V1 (old)	Metrics V2 (new)
1M	548	609	548
10M	10600	11160	10585

[akoshelev]

I've evaluated the performance impact of periodic flush based on the number of metrics emitted. I run a sharded test on 1 billion entries where metrics were sent to the collector thread after 8192 counter increments. The performance impact was 50% added latency time (2 hours vs 1 hour). On the other hand, there was no difference in accuracy. I haven't profiled this code thoroughly, but this change was the only one so I attribute 100% of regression to the elevated flush.