Operator decomposition row cache cleanup improvements

[morrone]

In issue #1483, and the associated PRs, a timeout was introduced to clean up stale row caches from operators in decomposition.

Really, what would be better is to cleanup the row cache dynamically when either:

• An individual group associated with a metric set goes away, or
• A metric set goes away, triggering cleanup of all related row caches

The rationale is that some metric sets may remain live, but not updated when there are no updates to the data in the metric set. As long as the metric set remains live, we would like to retain the row cache. With only a timeout method, the timeouts need to be carefully coordinated by a human to get them right.

If we cleared the caches dynamically, no humans need to spend time figuring things out to make the cleanup work correctly (or just giving up and picking a very long timeout). Dynamic cache cleanup would just do the Right Thing without human intervention.