Start local-queue jobs "asyncio" way

[warownia1]

Instead of using workers pool, utilize the fact that green threads are lightweight and can easily be spawned and disposed. Refer to comments in #72 .

Performance tests running 1M dummy jobs (sleep(0.0)) with 8 parallel workers

• Using queue and workers pool: mem: 195.0 MB (<0.1 MB overhead) time: 6.8 s (~6.8 us overhead per job)
• Spawning new worker for each job immediately and blocking it with a semaphore: mem: 3.7GB (>3.5 GB overhead) time: 46 s (~46 us overhead per job)
• Using job dispatcher limiting the number of parallel workers spawned for each job: mem: 195.4 MB (<0.1 MB overhead) time: 25 s (~25 us overhead per job)

[warownia1]

The advantage of having one worker per job makes it easier to cancel running jobs. The worker may catch CancelledError, terminate the sub-process, clean-up and die instead of looping forever. Additionally, failed workers will not deplete the pool and master thread is not needed restarting workers is not needed anymore.