Detect and replace dead worker processes

[aschampion]

Discovered while diagnosing #66 and an investigating an OOMing worker (hence noticing #65)

WorkerPool.reap_dead_workers checks process.is_alive() and removes exited
workers. TaskWorkerPools.check_worker_health calls it each event loop
iteration and spawns replacements. Without this, workers that die silently
(e.g., SIGKILL, SystemExit, OOM) leave the server hanging forever since
no error is queued and no blocks are returned. Includes a previously
failing test.

LLM disclosure: while I diagnosed and debugged the root issue, LLMs were used to exclude other causes and draft fixes. But all code here has been human reviewed or created.

[pattonw]

Good solution to a class of errors that daisy did not handle.

I think this will further exacerbate issue #64 since we don't have any checks on increment_workers to limit the number of workers we start. But that is a separate issue that should have its own pull request.