When we close a task pool, we use a condition variable to wait for
workers to exit, if any workers are still running. The workers,
in turn, use the condition variable to notify that they've exited.
Unfortunately, the workers aren't using the lock (the rationale was
that it's all atomic ops so they don't need to), which means that
it's possible to have the following sequence:
Close(): obtain current live count
worker: decrement live count
worker: send broadcast to condition variable
Close(): wait on condition variable
To resolve this, we make the worker update also request the lock.
We add a simple reproducer for this. Note that simple doesn't mean
it fires completely reliably; on my laptop, the test causes a test
timeout about 60% of the time without the fix. If you add a short
delay between sampling the live count and waiting on the condition
variable, the deadlocks move from "60% chance of hitting it in
a million trials" to "nearly always".
