feature(tj): optimize AlphaZero with batch inference support

[tAnGjIa520]

This PR introduces a batch-optimized AlphaZero MCTS implementation in C++, achieving 2x
speedup compared to the standard sequential version.

Batch MCTS Inference: The core improvement is the get_next_actions_batch() function (line
207-415 in mcts_alphazero.cpp), which processes multiple game states simultaneously. Instead of running MCTS
simulations sequentially for each environment, we now:

1. Batch Root Expansion: Initialize multiple root nodes and expand them with a single batched neural network call
   (policy_value_func_batch), reducing GPU overhead
2. Parallel Simulation Phase: Run simulations for all environments simultaneously (line 280-369), collecting leaf
   nodes that need expansion
3. Batch Leaf Expansion: Group all leaf nodes from unfinished games and perform batched inference via
   _batch_expand_leaf_nodes() (line 162-205), minimizing individual network calls
4. Legal Action Caching: Cache legal actions at the environment level to avoid repeated Python calls during child
   selection, significantly reducing Python-C++ interface overhead

本 PR 引入了批量优化的 AlphaZero MCTS C++ 实现，相比标准顺序版本实现了 2 倍加速。

批量 MCTS 推理: 核心改进是 get_next_actions_batch() 函数（mcts_alphazero.cpp 第 207-415
行），可同时处理多个游戏状态。我们不再为每个环境顺序运行 MCTS 模拟，而是：

1. 批量根节点扩展: 初始化多个根节点并通过单次批量神经网络调用（policy_value_func_batch）进行扩展，减少 GPU 开销
2. 并行模拟阶段: 同时为所有环境运行模拟（第 280-369 行），收集需要扩展的叶节点
3. 批量叶节点扩展: 将所有未完成游戏的叶节点分组，通过 _batch_expand_leaf_nodes()（第 162-205
   行）执行批量推理，最小化单独的网络调用
4. 合法动作缓存: 在环境层级缓存合法动作，避免在子节点选择期间重复的 Python 调用，显著减少 Python-C++ 接口开销