Fix wrong parent when nesting perf metrics

[copybara-service]

Fix wrong parent when nesting perf metrics

The refer_inference is currenlty nested under actor_training. Even though it is logically triggered before actor_training, its completion callback might run after actor_training has started. If actor_training is an active group on the same device timeline, the tracing infrastructure would incorrectly nest refer_inference under actor_training. Fixing the parent group to be at initiation time.

verified by manual launch.
https://screenshot.googleplex.com/5QmxLA7ZeTt8axQ

I0209 20:30:24.072806 131196942059584 export.py:419] Timeline: tpu0
I0209 20:30:24.072829 131196942059584 export.py:420] - root (0.000000, inf)
  - global_step (1.362577, 122.438728)
    - mini_batch_step (1.362613, 122.438719)
      - micro_batch_steps (1.362624, 122.438711)
        - rollout (1.392425, 35.577381)
        - actor_training (50.802475, 122.438482)
          - peft_train_step (122.164713, 122.437401)
        - refer_inference (35.659063, 50.838434)