Keep returns separate from advantages in GRPO

[ebronstein]

The GRPO advantage computation returns the normalized scores for both the advantages and returns even though it represents the advantages. This PR makes compute_grpo_outcome_advantage return the sequence-level sum of rewards for the returns, separately from the scores (i.e., advantages).

[gemini-code-assist]

Code Review

This pull request correctly separates the calculation of advantages and returns in compute_grpo_outcome_advantage. Previously, the function returned the computed advantages for both values. With this change, it now correctly returns the original sequence-level rewards as returns.

However, there is a critical issue with the shape of the returned returns tensor, which is currently 1D (batch_size,) instead of the expected 2D (batch_size, seq_len). I've left a specific comment with a suggested fix to address this.

Additionally, the unit tests for compute_grpo_outcome_advantage in skyrl-train/tests/cpu/utils/test_ppo_utils.py have not been updated to reflect these changes. The existing tests contain assertions that are now incorrect (e.g., asserting that advantages and returns are equal), and they would fail to catch the shape mismatch. Please update these tests to validate the new behavior.