Fix symmetric-run teardown errors on Slurm when processes exit as zombies

[ssam18]

When running ray symmetric-run on Slurm, Ray processes exit as zombies during teardown because ray stop is not their parent process and psutil.wait_procs() can't reap them via os.waitpid() that causes a false "Stopped 0 out of N" warning even though the job completed cleanly. Separately, worker nodes were always hitting a CalledProcessError on normal shutdown because ray start --block exits non-zero when GCS goes down, which is expected behavior and not an error. Both issues are fixed: zombies are now counted as pre-stopped before the signal loop, and check=True is removed from the worker's blocking ray start call. Fixes #62390.

[gemini-code-assist]

Code Review

This pull request updates the process termination logic to include zombie processes and modifies worker node startup by removing strict error checking. Review feedback highlights that the zombie process counting is over-inclusive and should be filtered to Ray-specific processes. Additionally, removing the error check in the worker startup command may silently ignore failures, suggesting a need for explicit return code verification or better documentation.

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{Reviewed by Cursor Bugbot for commit 2154977. Configure here.}