fix: preserve BSHD layout in PPO CP advantages

[guapisolo]

Summary

Preserve BSHD token ownership while computing PPO advantages with context parallelism.

Symptom & Reproduction

• Symptom: BSHD PPO with CP > 1 returned response-aligned outputs in THD token order.
• Reproduction: pytest -q -p no:cacheprovider tests/fast/backends/training_utils/test_ppo_cp_advantages.py::test_ppo_bshd_cp_uses_padded_layout_metadata

Root Cause

1. compute_advantages did not forward qkv_format or max_seq_lens.
2. get_advantages_and_returns_batch used THD defaults for gather and re-slice.

Fix

Propagate max_seq_lens from rollout data and pair it with args.qkv_format throughout PPO GAE. Both CP gather and re-slice now call the existing layout-aware helpers with identical per-sample metadata, while THD and CP=1 keep their previous behavior.

Verification

• New / updated test: test_ppo_bshd_cp_uses_padded_layout_metadata compares reconstructed CP=2 BSHD outputs with CP=1 baselines.
• Existing test suite: tests/fast/backends/training_utils/test_ppo_cp_advantages.py passes all three cases.

Review Focus

• Scrutinize metadata propagation from compute_advantages_and_returns to get_advantages_and_returns_batch.
• Scrutinize per-sample max_seq_lens pairing in both CP gather and re-slice.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.