fix(eval): abort all ranks on per-rank exception instead of deadlocking

[Luodian]

Summary

When torchrun / accelerate launches multiple ranks and one rank's evaluate() raises, today's handler in cli_evaluate logs the error, appends None, and lets the rank return normally. The other ranks keep going and block on the next collective (gather_object, barrier, etc.) on NCCL until the launcher's wall-clock timeout tears the job down — which on a SLURM cluster can waste hours of GPU time for a failure that is visible in rank 0's log at second zero.

Change

If torch.distributed is initialized when the outer exception handler fires, destroy the process group and sys.exit(1) so the launcher's elastic supervisor sees a non-zero rank and tears down the rest of the world immediately.

results_list.append(None)
+ if torch.distributed.is_available() and torch.distributed.is_initialized():
+     try:
+         torch.distributed.destroy_process_group()
+     except Exception:
+         pass
+     sys.exit(1)

Non-goals / scope

• No behavior change for single-process runs — the is_initialized() check gates the new path entirely.
• Does not try to distinguish retriable from fatal errors. The outer handler only fires after the inner eval loop already gave up.

Test plan

• Multi-rank run where rank N raises manually — verify all ranks exit within seconds instead of hitting wall-clock.
• Single-process run — verify no change in happy path.