Fix: Apply loss mask to KL in REINFORCE++ returns calculation

[kaysonyu]

Summary

Fixes #1370 - Apply loss mask to KL divergence before computing returns in get_reinforce_plus_plus_returns.

Problem

In REINFORCE++ advantage estimation, the KL divergence values were not masked before computing returns. This caused environment-generated tokens (e.g., tool_response in agentic RL scenarios) to incorrectly contribute to the KL penalty that propagates through the cumulative return calculation.

Solution

Apply loss_mask to zero out KL values at masked positions before computing returns:

  # Before:
  token_level_rewards = -kl_coef * full_kl_response

  # After:
  masked_kl = full_kl_response * full_mask
  token_level_rewards = -kl_coef * masked_kl