fix: clear stale error.pb at the start of every clustered attempt

[AdilFayyaz]

Motivation

For clustered (JobSet) tasks the SDK writes error.pb only on what it believes is the terminal
restart attempt, using JOBSET_RESTART_ATTEMPT >= JOBSET_MAX_RESTARTS. That guess mirrors the
JobSet's Status.Restarts, which also counts free host-maintenance restarts
(RestartJobSetAndIgnoreMaxRestarts, flyteorg/flyte#7947), while the budget is charged from
RestartsCountTowardsMax, which pods never see. So after a free restart the guess fires early, a
genuine failure writes error.pb on a non-terminal attempt, and because the executor reads
error.pb before outputs.pb, a later successful attempt is still reported as a failure.

Summary

• Add io.clear_stale_clustered_error(), which removes a leftover error.pb from the attempt's
  output prefix. It runs on every clustered rank-0 start, not just restarted ones, since the
  restart counter is exactly what proved untrustworthy. Only rank-0 deletes, restart attempts are
  strictly sequential, and a failed delete only restores today's behavior.
• Await it at the top of load_and_run_task, ahead of the code-bundle download and of all three
  upload_error call sites.
• Keep the terminal-attempt gate as an optimization that avoids needless writes, and rewrite its
  docstring: the guess can be early but never late, and the cleanup is what guarantees correctness.
• Rewrite examples/clustered/ddp_train_restart.py into a regression test that reproduces the
  stale file with no backend dependency, using max_restarts=0 plus one in-pod torchrun restart.

Test Plan

• uv run pytest tests/flyte/clustered tests/flyte/internal/runtime tests/flyte/internal/bin -q
  covers the async and sync delete branches, absent-file and non-rank-0 and non-clustered no-ops,
  soft failure on both the existence check and the delete, and that the cleanup is awaited before
  the task load, the load-failure error upload, and the task body. 364 passed.
• make lint and make mypy are clean.
• End to end on dogfood: make dist && uv run python examples/clustered/ddp_train_restart.py.
  Expect SUCCEEDED. flyte get logs <run> should show attempt 0 uploading .../a0/0/error.pb,
  then the restarted rank-0 logging Removed stale ... error.pb (JOBSET_RESTART_ATTEMPT=0, TORCHELASTIC_RESTART_COUNT=1), then training. Verified as run udvvrchxvfqpnwjvbfwp.
  Reverting the entrypoints.py hook makes the same run end FAILED with the attempt-0 error.