Align Rust action layout and self-play config with Python

[alejandromarcu]

by Codex. Trying to fix the difference between self-play in Python and Rust. This doesn't fix it but seems like legit bugs

Summary

• fix the policy/action layout comments, mask copying, and encoder conversions to match the Python [moves | vertical | horizontal] layout and keep tests aligned
• teach self-play configuration to merge overrides without clobbering defaults, compute the default noise-alpha using board size/walls, and pass board metadata when instantiating agents
• add Python-backed consistency tests for action encoding/action masks and document new experimental configs
  Testing