[codex] Add GPU kinodynamic graph-neural MPPI demo

[rsasaki0109]

Summary

• Add a CUDA kinodynamic graph-neural MPPI demo with Ackermann-like speed/steering rollouts.
• Compare a learned experience-graph guided local planner against a goal-only nonholonomic MPPI baseline.
• Register the gpu_kinodynamic_graph_neural_mppi CMake target and document the new GIF/benchmark in the README.

Validation

• git diff --check
• cmake --build build --target gpu_kinodynamic_graph_neural_mppi -j$(nproc)
• ./bin/gpu_kinodynamic_graph_neural_mppi
• ctest --test-dir build --output-on-failure --label-regex 'cpu|python|gpu' -j$(nproc)\n\n## Demo Result\n- 32768 nonholonomic speed/steering rollouts x H=72 x guided/unguided batches\n- Cost: 1516.74 -> 851.11 (-43.9%)\n- Terminal error: 5.11 -> 0.88 (-82.8%)\n- Route error: 1.530 -> 0.252\n- GPU: 33.154 ms, 49.9x vs CPU equivalent kinodynamic rollout evaluation