Concurrency fixes

[lihaoyi]

• Swap from ForkJoinPool to ThreadPoolExecutor

  • ForkJoinPool does this weird thing where fork-join-tasks can be re-entrant at join points, resulting in weird scenarios where a mill-task that hits a yield point (e.g. inside Zinc / parallel-collections / FJP) can start running a second mill-task even before the first has finished, violating all sorts of invariants (# of running tasks exceeds --jobs, FixSizeCache semaphores get taken twice by the same thread, all sorts of craziness)
  • We replace ForkJoinPool#ManagedBlocker with our own manual logic increasing and decreasing the ThreadPoolExecutors maximumPoolSize and corePoolSize in our blocking{...} wrapper

• We need to Thread#interrupt() the promptUpdaterThread thread when we close the PromptLogger, so we don't need to wait the promptUpdateInterval (0.1ms for interactive, 60s for non-interactive) before exiting

This should fix some of the flakiness we've been seeing in master, that seems to have started from 05bef7e (just eyeballing the CI history), and blocking our re-bootstrapping in #3637