[codex] Add MPPI diagnostics follow-up tooling

[rsasaki0109]

Summary

Adds the post-v0.1.0 MPPI follow-up tooling and documentation:

• checked-in extended controller scenario results for double_gap and moving_crossing
• renderer for cuda_mppi_controller diagnostics CSV files
• bag / real-data evaluation harness for Nav2 CUDA MPPI runs
• registration external benchmark stress scenarios: low_overlap, outlier_partial, and large_offset
• no-GPU ROS2 CI runtime smoke tests for plugin loading and parameter validation
• Python DLPack costmap example for CUDA PyTorch / CuPy users
• plan/docs/readme updates, including SOPPI detour guidance to avoid blind same-mechanism tuning

Validation

• python3 -m py_compile scripts/render_cuda_mppi_extended_scenarios.py scripts/render_cuda_mppi_diagnostics.py scripts/run_cuda_mppi_bag_eval.py scripts/benchmark_registration_external.py examples/python/mppi_dlpack_costmap.py python/src/cudarobotics/__init__.py python/tests/test_import.py
• git diff --check
• ros2 run cuda_mppi_controller plugin_load_test
• ros2 run cuda_mppi_controller parameter_validation_test
• python3 scripts/run_cuda_mppi_bag_eval.py --output-dir /tmp/cuda_mppi_bag_eval_smoke --controller-command 'sleep 5' --duration 0.2 --settle-seconds 0 --no-record --no-render-diagnostics
• locally ran controller_benchmark for double_gap cpu_gpu and moving_crossing quick, then rendered docs/results/cuda_mppi_extended_scenarios_2026-06-12.{csv,md}

Notes

The Python DLPack example requires an installed cudarobotics package and either CUDA PyTorch or CuPy at runtime. The registration stress benchmark also requires the same external benchmark environment as the existing registration report.