Add planner-oriented atomic skill benchmark

[yuecideng]

Description

This PR adds a shared planner benchmark framework for physics-backed Atomic Actions and implements the first Franka + PGI and cuRobo vertical slice proposed in #437.

The benchmark now:

• routes Atomic Actions through an adapter-owned MotionGenerator with MotionPolicy(strategy="motion_gen", planner="curobo");
• registers planners, robots, scenarios, atomic skills, and object kinds behind shared extension points;
• freezes robot, object, grasp, difficulty, and independent IK evidence in case manifests;
• separates planning, motion-validity, execution, and physical task success;
• reports planning and execution efficiency, controller tracking, path quality, object lift, memory, and a complete leaderboard in exactly three tables;
• provides a deterministic atomic_franka_pgi_curobo smoke suite covering MoveEndEffector and antipodal-grasp PickUp;
• preserves the existing free-space benchmark through the same runner.

Current Atomic Task scope is deliberately limited to batch size 1, an empty external cuRobo collision world, and the two initial skills. Additional planners opt in by exposing a MotionGenerator and the atomic_action capability.

Dependencies: cuRobo and CUDA are required to execute the new suite. No dependency manifests are changed.

Refs #437

Type of change

• Bug fix (non-breaking change which fixes an issue)
• Enhancement (non-breaking change which improves an existing functionality)
• New feature (non-breaking change which adds functionality)
• Breaking change (existing functionality will not work without user modification)
• Documentation update

Screenshots

Not applicable. The feature produces JSON/JSONL artifacts and a Markdown benchmark report.

Validation

• black --check --diff --color ./ — 671 files unchanged
• pytest -q tests/benchmark/motion_generation — 30 passed, 1 skipped
• python -m compileall -q scripts/benchmark/motion_generation tests/benchmark/motion_generation/test_atomic_task_benchmark.py
• python -m scripts.benchmark.motion_generation.run_benchmark --suite atomic_franka_pgi_curobo --device cuda — MoveEndEffector and PickUp both reached 100% planning, motion, execution, and task success; PickUp lifted the object by approximately 0.1594 m

Checklist

• I have run the black . command to format the code base.
• I have made corresponding changes to the documentation
• I have added tests that prove my feature works
• Dependencies have been updated, if applicable. No dependency changes are required.