Fix canary validation artifact replication (#4173)

## Summary
- restore canary `tracker_metrics.jsonl` emission by setting
`replicate_path=this_output_path()` on the canary W&B tracker
- add a regression test that reloads the canary step and asserts metrics
replication stays configured
- fix the TPU canary post-run validation failure seen in scheduled runs
such as 23581948912, where training succeeded but metrics validation
failed because the artifact was never written

## Testing
- `./infra/pre-commit.py experiments/ferries/canary_ferry.py
tests/test_validate_canary_metrics.py`
- `uv run --with pytest --with pytest-timeout --with pytest-xdist python
-m pytest tests/test_validate_canary_metrics.py`

---------

Co-authored-by: yoblin <268258002+yoblin@users.noreply.github.com>

Author: 2 people

experiments/ferries/canary_ferry.py

@@ -138,6 +138,7 @@ def _build_step_from_env() -> ExecutorStep:
                 tags=wandb_tags,
                 group=wandb_group,
                 name=None,
+                replicate_path=this_output_path(),
             ),
             optimizer=versioned(CANARY_OPTIMIZER),
             grug_trainer=versioned(CANARY_TRAINER),