[codex] Add MPPI path-angle critic

[rsasaki0109]

Summary

• add a GPU MPPI path-angle critic controlled by path_angle_weight (default 0.25)
• expose the parameter through C++/Python core headers, Python bindings, Nav2 dynamic params, validation, and example configs
• add controller_benchmark ... path_angle plus checked-in CSV/Markdown results, and refresh the ESDF report with the new default critic enabled
• update README/docs/site/plan references for the new critic and benchmark report

Notes

The rollout kernel computes a local path tangent around the follow point and adds a light yaw-error stage cost. Reverse samples compare against path_tangent + pi; the existing backward-motion critic still decides whether reverse motion is worth using.

A stronger path_angle_weight=2.0 over-selected stationary straight-heading rollouts in the standalone wall-gap smoke. The checked-in default is tuned to 0.25, which keeps the straight/corridor smoke cells successful and improves the corrected u_turn K8192 time-to-goal from 44.75 s to 39.9 s in the local benchmark.

Validation

• source /opt/ros/humble/setup.bash && colcon build --packages-select cuda_mppi_controller --cmake-args -DCMAKE_BUILD_TYPE=Release
• source /opt/ros/humble/setup.bash && source install/setup.bash && ros2 run cuda_mppi_controller parameter_validation_test
• source /opt/ros/humble/setup.bash && source install/setup.bash && ros2 run cuda_mppi_controller mppi_gpu_standalone 2048
• source /opt/ros/humble/setup.bash && source install/setup.bash && ros2 run cuda_mppi_controller controller_benchmark /tmp/mppi_path_angle_u_turn u_turn
• source /opt/ros/humble/setup.bash && source install/setup.bash && ros2 run cuda_mppi_controller controller_benchmark /tmp/mppi_path_angle_esdf esdf
• source /opt/ros/humble/setup.bash && source install/setup.bash && ros2 run cuda_mppi_controller controller_benchmark /tmp/mppi_path_angle_bench path_angle
• python3 -m py_compile scripts/render_cuda_mppi_esdf_benchmark.py python/src/cudarobotics/__init__.py
• git diff --check
• forbidden hardware-string scan: no matches

Python native extension smoke was not rerun locally because scikit-build-core and nanobind are not installed in this environment.