fix: abort process groups on rank-local training failure

[maocheng23]

Motivation

Stack 2/5 replacing draft #766. When one rank dies inside a run (a data or capture error) while its peers are blocked in a CUDA/FSDP collective, the collective destroy_process_group in the CLI's finally hangs in NCCL communicator destruction: the originating traceback never surfaces and torchrun cannot reap the job. Colocated online training makes rank-local capture failures a realistic event, but this fix applies to every distributed topology.

Modifications

• destroy_distributed(abort=False): on the exceptional path, call the non-collective ProcessGroup.abort() on every distinct cached group, including the default group, so elastic surfaces the real exception. The success path keeps the existing collective destroy; torch builds without ProcessGroup.abort fall back to destroy.
• The CLI _train marks the exceptional path with a failed flag and passes abort=failed to teardown.

Related Issues

Splits #766. Stack: #1 capture-rows ← #2 (this) ← #3 rank0-tracker ← #4 colocated-core ← #5 hybrid-shard.

Accuracy Test

• New test_failure_teardown_aborts_each_distinct_process_group verifies abort is called once per distinct group and collective destroy is not used on the failure path.
• Existing teardown test keeps covering the success path.

Checklist

• Format your code according to the Code Formatting with Pre-Commit (black --check and isort --check-only).
• Add unit tests.
• Update documentation as needed (none needed).