Don't lose worker startup exception when its process exit is observed first (flaky test_local_cluster_redundant_kwarg)

[crusaderky]

Fix race condition in Nanny when a worker fails to start.

Follow-up to #9313; fixes the remaining flakiness in test_local_cluster_redundant_kwarg (example failure).

In-depth commentary (AI-generated)

Root cause

When the worker process fails to start, the child sends the startup exception to the Nanny through init_result_q, flushes the queue, and terminates. WorkerProcess._wait_until_connected polls that queue every 50ms and, in each iteration, first bails out with None if the status is no longer Status.starting. If mark_stopped() (fired by the process exit callback) flipped the status during one of the 50ms sleeps, before the poll picked up the message, the exception was lost: instantiate() returned Status.stopped and Nanny.start_unsafe raised a bare RuntimeError("Nanny failed to start worker process") instead of chaining from the worker's original exception, breaking the test's raises_with_cause(RuntimeError, None, TypeError, "unexpected keyword argument") expectation.

Fix

• Poll init_result_q one last time after observing that the process stopped. The child always flushes the queue before exiting (close() + join_thread()), so by the time the exit callback has run, the message — if any — is guaranteed to be readable.
• Hold a local reference to init_result_q in WorkerProcess.start(), since mark_stopped() releases self.init_result_q.
• Don't overwrite the status with Status.running if the process already died: mark_stopped() set it to Status.stopped and released the queues, and overwriting it would make the next kill() crash on the released queues, leaving the Nanny stuck in Status.closing (same failure mode described in Fix Nanny race conditions when worker fails to start (flaky test_failure_during_worker_initialization) #9313).