fix: preserve raw PPO reference KL

[hkxxxxx]

This fixes an in-place mutation bug in Slime's PPO reference-KL reward shaping path.

Before this change, the raw reference-policy KL tensor was multiplied by kl_coef in place before Reef could consume it for adaptive-KL telemetry. The change preserves the raw KL tensor and uses a separate reward-shaping tensor.

Validation:

• tests/test_ppo_reference_kl.py: 1 passed
• No model, recipe, or default behavior changes.

[hkxxxxx]

This fix is already included upstream in #2114
(commit 045310b, "fix(ppo): preserve raw KL so rollout/kl logging is correct").

The current THUDM/main implementation already preserves raw KL,
so this PR is no longer needed.