Thread-safe task store with worker dispatch and requeue

Build a task store that many threads use at once. Tasks are queued, running or completed, and each worker runs at most one task at a time.

• add(task) queues a new task.
• dispatch(worker) hands the worker the oldest queued task, or nothing when none is queued.
• complete(worker, task) marks it done and frees the worker; completing the same task twice is harmless.
• fail(worker, task) frees the worker and puts the task back in the queue at its original position, not at the end; release(worker) does the same for whatever the worker is running.

Make every call atomic and explain how you keep the queue order and avoid races.