[codex] Add GPU multi-agent graph-neural MPPI demo

[rsasaki0109]

Summary

• Add a CUDA multi-agent graph-neural MPPI demo for 48 robots crossing a shared interaction zone.
• Compare independent per-robot MPPI against coordinated graph-message MPPI with peer-risk-aware rollout selection.
• Register the gpu_multiagent_graph_neural_mppi target and document the new GIF/benchmark in the README.

Validation

• git diff --check
• cmake --build build --target gpu_multiagent_graph_neural_mppi -j$(nproc)
• ./bin/gpu_multiagent_graph_neural_mppi
• ctest --test-dir build --output-on-failure --label-regex 'cpu|python|gpu' -j$(nproc)\n\n## Demo Result\n- 48 robots x 768 rollouts x H=72 x independent/coordinated modes\n- Cross-route collisions: 518 -> 261 (-49.6%)\n- Social risk: 3.544 -> 2.588 (-27.0%)\n- Minimum separation: -0.420 -> -0.398 (+0.022 m)\n- Reach basin: 48/48 -> 36/48\n- GPU rollout eval: 5.660 ms, 3139.6x vs CPU equivalent rollout evaluation