docs: clarify PPO entropy metrics in PPO trainer docs

[biefan]

Summary

Clarify the difference between objective/entropy and policy/entropy_avg in the PPO trainer docs.

What changed

• Updated the objective/entropy description to match the rollout-time computation ((-logprobs).sum(1).mean()).
• Updated the policy/entropy_avg description to match the optimization-time entropy computed from logits.
• Added a short note explaining why these two metrics are expected to differ.

Why

Issue #2023 points out that the two entropy metrics had very similar wording, which made interpretation difficult when debugging PPO runs.

Fixes #2023

---

Note

Low Risk
Low risk documentation-only change that updates metric wording and adds a brief clarification note; no runtime or API behavior is modified.

Overview
Clarifies the PPO trainer metric docs by rewriting the descriptions of objective/entropy (rollout-time proxy computed from -logprobs) and policy/entropy_avg (optimization-time categorical entropy computed from logits).

Adds an explicit note explaining that these metrics are measured at different phases (rollouts vs. PPO optimization) and therefore are expected to differ.

^{Written by Cursor Bugbot for commit 8367f1e. This will update automatically on new commits. Configure here.}