fix: delete stale error.pb before uploading outputs on JobSet restart

[AdilFayyaz]

Motivation

When a clustered (JobSet) task fails, it writes an error.pb to the output path. If a later restart attempt succeeds, that stale error file is left behind and the action can be misreported as failed despite a successful run.

Summary

• Add _clear_stale_clustered_error_if_needed() to upload_outputs: on restart attempts (JOBSET_RESTART_ATTEMPT > 0), delete a leftover error.pb from the output path before writing outputs.
• Add _delete_path() helper that deletes via the underlying fsspec filesystem, handling both async (_rm_file/_rm) and sync (rm_file/rm) backends.
• Delete is best-effort: missing files are debug-logged, other failures are downgraded to a warning so output upload still proceeds.
• Attempt 0 is a no-op (no exists/filesystem calls), keeping the non-restart hot path unchanged.

Test Plan

• New tests/flyte/clustered/test_runtime_io_cleanup.py covers:
• restart attempt > 0 deletes the stale error then uploads outputs (correct call order),
• attempt 0 skips the existence check / delete entirely,
• a delete failure is soft-failed (warning logged, outputs still uploaded).
• Run: uv run pytest tests/flyte/clustered/test_runtime_io_cleanup.py -v