Ensure task start is interrupted, if its assigned group or the entire pool is cancelled

[daniil-berg]

Right now the _start_task method may be subject to a context switch, if the pool size is limited and there is no room for the new task at this point:

asyncio-taskpool/src/asyncio_taskpool/pool.py

Line 395 in 27eb3bf

await self._enough_room.acquire()

Since the _start_task coroutine is itself scheduled in a task from TaskPool.apply, TaskPool._map, and SimpleTaskPool.start (by proxy of TaskPool._apply_spawner, TaskPool._arg_consumer, and SimpleTaskPool._start_num respectively), this can introduce a race condition:

The _start_task coroutine may still be awaiting the semaphore for enough room, while a cancel was already issued for the group it is assigned to or the entire pool. Cancellation right now only affects tasks in the _tasks_running dictionary. This means the _start_task call may still spawn a new task and add it to _tasks_running after it was supposedly cancelled.