perf: Improve performance of stress relief

[kentquirk]

Which problem is this PR solving?

In testing stress relief, I found that it could turn off too easily, and it might then cause refinery to crash.

Short description of the changes

• Don't let stress relief turn off until stress levels are below the deactivation level for 10 consecutive seconds
• Move the calculation of stress relief into the periodic stress check, rather than doing it on every span. Now each span check needs only a read lock.
• Tweak the math for sigmoid (it had a little slop that was bugging me).