[codex] Add reproducible minimal PPO WSL workflow

[HC-Seaple]

What changed

• Add a self-contained continuous-action PPO trainer with vectorized rollouts, GAE, clipped updates, checkpointing, and deterministic evaluation.
• Add Windows/WSL setup and launch scripts for the Linux-native Raylib build.
• Add generic WOMD JSON-to-map preparation without committing datasets or generated binaries.
• Add native third-person checkpoint visualization and JSON metrics.
• Ensure complete renderer frames are written to ffmpeg.
• Document clone, setup, map preparation, training, visualization, and handoff.

Validation

• Python scripts pass python -m py_compile.
• WSL launchers pass bash -n.
• The staged change set passes git diff --check.
• The end-to-end workflow was previously exercised in WSL with a 10,112-step checkpoint and 92-frame native render.

Current limitation

This is a smoke-test training architecture. Reward shaping still needs route-progress reward, reverse-motion penalties, and stronger collision/off-road costs before scaling.

[eugenevinitsky]

Hi! Thanks for the PR. We can review and discuss this once it's off draft status but a quick thing to mention is that this probably needs a slightly different folder structure since otherwise it pollutes the script folder with a lot of unstructured code