fix: move Unity lab into PPO chapter

Author: sanbuphy

docs/chapter10_ppo/ppo-game-benchmark.md

@@ -195,6 +195,8 @@ Mario 是经典项目，但**奖励很容易写歪**。只奖励向右走，策
 
 SoccerTwos 是 2v2 足球，引入了一个全新维度：**多智能体**。PPO 在这里需要和 **self-play** 配合使用——智能体与历史版本的自己对战，通过 **ELO 评分**追踪进步。**Unity ML-Agents 内置 PPO** 是最直接的实现路径，batch_size **2048**、buffer_size **20480**、学习率 **$3 \times 10^{-4}$**、熵系数 **0.005**。
 
+<OnlineTraining studios="unity" compact />
+
 训练过程能看到明显的**策略演化**。初期所有智能体追着球跑，场面混乱；ELO 从初始 **1200** 慢慢爬到约 **1600** 后，开始出现简单的**前锋-后卫角色分化**。**SAC** 在这个环境上表现差很多，基本停留在 **1200-1250**，说明 **PPO 的稳定性对多智能体 competitive 场景很重要**。
 
 **课程学习**的效果很显著。设计良好的课程（逐步调整初始位置、球速）能让 PPO 在 **250k iterations** 内平均奖励超过 **1.8**，节省约 **40%** 训练时间；但设计不好的课程反而降低性能（**0.35 vs 0.45**），这提醒我们**课程设计本身需要调优**。

docs/en/chapter10_ppo/ppo-game-benchmark.md

@@ -195,6 +195,8 @@ PPO often fails on unfamiliar terrain hazards rather than because the low-level
 
 SoccerTwos is 2v2 soccer, so it introduces **multi-agent learning**. PPO is usually paired with **self-play**: agents play against historical versions of themselves, and progress is tracked with ELO. Unity ML-Agents provides the most direct PPO implementation path.
 
+<OnlineTraining studios="unity" compact />
+
 The training process often shows strategy evolution. Early agents chase the ball chaotically. As ELO rises from **1200** toward **1600**, simple attacker-defender role separation can appear. SAC tends to perform much worse in this setting, which highlights PPO's stability in competitive multi-agent training.
 
 Curriculum learning matters. Gradually changing start positions and ball speed can reduce training time, but a poorly designed curriculum can hurt final performance. Multi-agent reward design is also delicate: rewarding only goals is sparse; over-rewarding distance to the ball makes every agent crowd the same location.

docs/en/preface/env-setup.md

@@ -135,8 +135,6 @@ pip install numpy scipy matplotlib seaborn pandas
 pip install tqdm tensorboard wandb
 ```
 
-<OnlineTraining studios="gymnasium" compact />
-
 ## Simulator Environments
 
 Different chapters use different simulators. Install them as needed.
@@ -211,8 +209,6 @@ Using ML-Agents also requires downloading or building Unity environments (`.exe`
 The unique value of ML-Agents is **3D spatial reasoning**: Atari uses 2D pixels, CartPole uses low-dimensional vectors, while ML-Agents provides a complete 3D physics environment (gravity, collisions, occlusion). If your research involves visual navigation, spatial reasoning, or multi-agent 3D coordination, ML-Agents is a strong complement to Gymnasium/PyBullet.
 :::
 
-<OnlineTraining studios="unity" compact />
-
 ## LLM Training Frameworks
 
 The second half of the course (Chapters 7-10) involves large model alignment training.

docs/preface/env-setup.md

@@ -137,8 +137,6 @@ pip install numpy scipy matplotlib seaborn pandas
 pip install tqdm tensorboard wandb
 ```
 
-<OnlineTraining studios="gymnasium" compact />
-
 ## 仿真环境安装
 
 不同章节用到不同的仿真环境，可以按需安装。
@@ -213,8 +211,6 @@ python -c "from mlagents_envs.environment import UnityEnvironment; print('ML-Age
 ML-Agents 的独特价值在于**3D 空间推理**：Atari 是 2D 像素，CartPole 是低维向量，而 ML-Agents 提供完整的 3D 物理环境（重力、碰撞、遮挡）。如果你的研究涉及视觉导航、空间推理或多智能体 3D 协作，ML-Agents 是 Gymnasium/PyBullet 之外的有力补充。
 :::
 
-<OnlineTraining studios="unity" compact />
-
 ## LLM 训练框架
 
 课程后半部分（Ch7-Ch10）涉及大模型对齐训练。