There's apparently a few places where we do not call ->finalize
or ->finish and leave dangling limiter slots occupied. I can't
reproduce this easily, so it's likely in error-handling paths.
I already made ->finalize idempotent when switching to awaitpid
since I wanted to rely entirely on DESTROY. However, DESTROY
doesn't always fire soon enough (and the client has already seen
a response), but using DESTROY as a fallback seems reasonable..
This does the minimum to ensure the limiter is freed up on
process exit, but ensuring a finish/finalize call always happens
is the goal.