fix(rlix): release train GPUs after sync failure

[TianyeGGBond]

What

Wraps sync_base_weights_to_active() in _after_training() with try/finally so the actor_train allocation is released even if weight sync fails after the train actor has already offloaded.

Why

Before:

sync_base_weights_to_active()
_notify_release_cluster_gpus()

If sync raised, _notify_release_cluster_gpus() was skipped and the scheduler could keep actor_train marked as allocated until shutdown_hard.

In multi-pipeline runs, that can block peer pipelines from stepping.

After

try:
    sync_base_weights_to_active()
finally:
    _notify_release_cluster_gpus()

The original sync exception still propagates. This only prevents leaking the scheduler allocation.

Tests

python -m pytest -q tests/test_miles_pipeline_after_training_cleanup.py

1 passed in 0.03s

Scope

Exception-safety hardening only. No happy-path behavior change.

[taoluo]

lgtm