fix: place PPO terminal rewards on the owning CP rank

[MrCapricornLiu]

PPO currently adds the scalar reward to CP rank 0. With zigzag context-parallel slicing, padding can put the final response prediction on another rank. For example, total length 9 and response length 3 at CP=2 leave rank 0 with no response predictions, so indexing its last reward raises IndexError. If rank 0 owns earlier response predictions, the reward is instead attached to the wrong time step.

Use the existing CP token offsets to find the shard containing the final response token. CP=1 retains the existing behavior, and the per-token KL reward remains unchanged.

Validation: the added CPU/Gloo regression runs the real advantage/return calculation at CP=1, 2 and 4, comparing five sequence layouts with an independent unsharded GAE recurrence. The original code fails the CP=2/4 cases; the updated test module passes all 4 tests. The CP utility, discounted-return and CP whitening modules also pass (56 tests). Changed-file pre-commit hooks pass.

The distributed tests use the repository's MPU stub for rank/group metadata and real Gloo collectives. They do not run a Megatron model, NCCL, or a full PPO training job.

OMP_NUM_THREADS=1 PYTHONPATH=.:tests pytest -q tests/test_ppo_kl_metric.py
OMP_NUM_THREADS=1 PYTHONPATH=.:tests pytest -q tests/test_cp_utils.py tests/test_discounted_returns.py tests/test_advantage_whiten_cp.py