Port GPU headless renderer to gsp_v0

[julianh65]

Summary

• Port GPU-accelerated headless rendering from fe68642 onto gsp_v0
• Add optional EGL headless path with Xvfb/Mesa fallback
• Add static road-geometry cache and PBO/writev frame readback
• Keep eval video retention configurable via eval.save_render_videos
• Use full-map headless FULL_SIM_STATE rendering to match the upstream renderer PR look

Validation

• git diff --check
• NO_TRAIN=1 .venv-renderer-run/bin/python setup.py build_ext --inplace --force
• Small BC renderer run: 3 scenarios, 50 frames each, logged to wandb: https://wandb.ai/emerge_/pufferdrive-renderer-validation/runs/2ajrdmgk

Notes

• Local machine does not have EGL/egl.h, so the small run validates the Xvfb/Mesa fallback path plus the full-map renderer/road cache/recorder plumbing.
• EGL-specific path should be validated on a machine with EGL dev headers installed.

[julianh65]

Updated with commit 073eb0af1 to match the upstream headless trajectory breadcrumbs (DrawPoint3D instead of sphere traces). Rebuilt locally and ran a BC-policy visual validation logging both full-map and perspective videos: https://wandb.ai/emerge_/pufferdrive-renderer-validation/runs/3uq0yj3p

Note: this desktop still lacks EGL headers, so the local validation uses the Xvfb/Mesa fallback rather than the actual EGL GPU path.

[julianh65]

Updated with d21d97c57 to match the 2D clean-eval render style from the referenced run: fixed 1920x1080 headless output, centered orthographic full-map camera, green filled policy-agent boxes, and no headless text overlays. New 2D validation run: https://wandb.ai/emerge_/pufferdrive-renderer-validation/runs/r1jpl38m

[julianh65]

Updated with 89fe3f65f to copy the clean-eval visual styling for the remaining mismatches: active policy vehicles use filled Raylib GREEN, expert/static vehicles use GOLD, and goals use LIME markers with faint LIME radius circles. Corrected validation videos: https://wandb.ai/emerge_/pufferdrive-renderer-validation/runs/qob2btjg