DeepSpeed releases process-group caches before CPU/Gloo worker shutdown

The official source reports this update: nightly-last-green: Release process-group caches before CPU/Gloo worker shutdown (#8846). Problem After successful distributed teardown, DeepSpeed could retain references to destroyed process groups in its group caches. Reinitialization could then reuse a stale TP group. In the CPU/Gloo AutoEP test path, completed engines could also remain in hook cycles until interpreter finalization; native Linux reproduced a SIGABRT while a Gloo worker released Python tensor state during finalization.