[fix][train] Shut down DataLoader workers on exit

[bvolpato]

Summary

StatefulDataLoader owns worker processes through its iterator. RL and SFT entrypoints only cleaned up dataloaders on successful runs, so training exceptions could leave train or evaluation workers alive.

Changes

• Add a guarded, idempotent shutdown helper for torchdata 0.11.
• Shut down RL train and evaluation dataloaders on success and failure.
• Move SFT shutdown into the entrypoint's finally block and close every train and evaluation dataloader.
• Log cleanup failures without masking the original training exception.

Normal Python unwinding is covered. Abrupt termination such as SIGKILL remains outside scope.

Testing

• Focused lifecycle tests (123 passed)
• Full prior SkyRL Train CPU suite (1467 passed, 5 skipped, 5 deselected)
• Tests start real train and evaluation workers, call shutdown twice, and verify every worker exits
• pre-commit run --all-files
• git diff --check upstream/main

[gemini-code-assist]

Code Review

This pull request ensures that trainer resources, specifically DataLoader worker processes, are properly shut down when training finishes or encounters an error. It introduces a shutdown_dataloader utility function to stop worker processes owned by a StatefulDataLoader, and integrates this into the shutdown methods of both SFTTrainer and Trainer. Additionally, finally blocks are added to the entry points to guarantee that shutdown is always executed. Feedback suggests improving the readability of the dataloader shutdown logic in SFTTrainer by separating the train and evaluation dataloader shutdown blocks instead of using an intermediate list.