metric: optimize Histogram.RecordValue (1.2% cpu on sysbench)

[tbg]

on current master, histogram observations account for >1% of cpu.¹
This may be due to the bucket count, though the profile only spends a portion of the time in the bucket search.
Sadly most of the time is spent in dependencies. Still, this is a sizeable chunk of CPU.

https://pprof.me/fcd1cbfe64d478dd1e6211970ff9107e/?profileType=profile%253Acpu%253Ananoseconds%253Acpu%253Ananoseconds%253Adelta&filter_by_function=Observe&search_string=Observe

Jira issue: CRDB-43623

Footnotes

1. https://github.com/cockroachlabs/perf-efficiency-team/pull/4/commits ↩

[tbg]

Took a quick look here and nothing jumps out. This is just a very very hot code path. It doesn't help that we maintain two histograms per histogram (one for windowing, one for the cumulative counters). We could save half the work if we updated only the most recent one, and then merged into the cumulative at each rotation interval - that would give a factor of ~0.5 here.