sql: recent regression in stopper quiescence time

[andreimatei]

#46832 seems to have somehow caused a regression for tests shutting down. Every now and then, shutting down a server seems to take ~15s, and the logs show:

W200403 19:36:34.837125 3038 sql/temporary_schema.go:430  [n1] error during schema cleanup, retrying: node unavailable; try another peer
W200403 19:36:36.577356 3038 sql/temporary_schema.go:430  [n1] error during schema cleanup, retrying: node unavailable; try another peer
W200403 19:36:41.068366 3038 sql/temporary_schema.go:430  [n1] error during schema cleanup, retrying: node unavailable; try another peer
W200403 19:36:49.157471 3038 sql/temporary_schema.go:430  [n1] error during schema cleanup, retrying: node unavailable; try another peer
W200403 19:36:49.157505 3038 sql/temporary_schema.go:551  [n1] failed to clean temp objects: node unavailable; try another peer
I200403 19:36:49.157515 3038 sql/temporary_schema.go:562  [n1] temporary object cleaner next scheduled to run at 2020-04-03 20:06:33.717927 +0000 UTC

I've bisected it pretty conclusively to that PR. I'm looking superficially at the PR though, and I can't tell what's wrong.
It seems to affect tests at random. The effect is big enough to cause the tests for the sql package to take a lot longer than they used to.

To repro, for example, you can do do

make testshort PKG='./pkg/sql' TESTFLAGS="-v --count=50" TESTS=TestSavepointMetric

One out of every 10 runs or so will be very slow.

The PR in question was backported to 20.1 too. I'm gonna mark it as a release blocker cause I find it pretty scary, but it might turn out to not be too bad.

[knz]

oh yeah I had noticed that but so far I've found it only affects TestServers. I did not see it with cockroach start nodes. I'll investigate more.

[RaduBerinde]

This is the same with #47047 which has been fixed in #47048.