Adaptive metrics probing periods

[ahg-g]

What would you like to be added:

Adaptive metrics probing periods. I can think of two options:

1. Either change the probing period depending on the observed qps: reduce it as qps increases, increase it as qps decreases within pre-defined bounds.
2. Explore a form of synchronous probing

Why is this needed:

Currently the EPP has a fixed probing period of 50ms, at higher qps this probing period may not be sufficient, leading to stale metrics and suboptimal scheduling decisions.

[liu-cong]

I discussed this in #678

[kfswain]

Issue Context: A PR that implements this would need to show improved benchmark results to better quantify the value.