Update normalizers after optimization

[NeoZng]

Summary

• keep observation-normalization statistics fixed throughout PPO and distillation optimization
• update actor, critic, student, and RND normalizers once from the flattened rollout observations after optimization
• use the observations that were actually stored and optimized instead of the next observations passed to process_env_step()

Problem

process_env_step() currently updates observation normalizers after every environment step. Actions and their old log probabilities are computed before that update, while PPO.update() recomputes log probabilities after the normalizer has changed. Consequently, the first policy ratio can differ from one even though no policy parameter update has occurred yet.

The per-step update also consumes the post-step observation (s_{t+1}), whereas the optimization batches contain the pre-step observations (s_t) stored in the rollout.

Implementation

• remove normalizer updates from PPO and distillation process_env_step()
• flatten the rollout time and environment dimensions and update normalizers after all optimization epochs
• perform the update before clearing storage so the next rollout uses the newly accumulated statistics
• apply the same timing to PPO actor/critic/RND normalizers and the distillation student normalizer
• add the contributor entry required by CONTRIBUTING.md

Tests

• pytest tests/algorithms/test_ppo.py tests/algorithms/test_distillation.py -q — 20 passed
• pre-commit run --all-files — all hooks passed

The new PPO regression test verifies that the first recomputed ratio remains one before optimization and that both normalizers are updated exactly once from stored rollout observations. The distillation regression test verifies the corresponding student-normalizer timing and data source.