benchmark: add cross-backend physics_step throughput coverage for genesis, isaacgym, and mujoco_warp

[TATP-233]

Background

We already have a grouped set of physics-step benchmark entrypoints under benchmark/:

• benchmark/benchmark_physics_step_mj_step.py
• benchmark/benchmark_physics_step_genesis.py
• benchmark/benchmark_physics_step_isaacgym.py
• benchmark/benchmark_physics_step_mujoco_warp.py

But we still lack a clear, repo-level benchmark target / support claim for comparing physics_step throughput across backends, especially for genesis, isaacgym, and mujoco_warp.

Goal

Add benchmark coverage and a stable reporting path for cross-backend physics_step throughput, including at least:

• genesis
• isaacgym
• mujoco_warp
• baseline MuJoCo path (benchmark_physics_step_mj_step.py)

Expected Scope

1. Standardize the measurement target as physics-step throughput / latency, not full env.step throughput.
2. Ensure the backend-specific benchmark scripts expose comparable output fields and compatible task / batch-size sweeps.
3. Save results under the shared benchmark/outputs/physics_step/ subtree so results are easy to compare and postprocess.
4. Make it easy to compare:
   • total throughput
   • per-env throughput
   • latency per benchmark call
5. Keep task naming aligned with current owner task ids (go1_joystick, go2_joystick, g1_joystick) and avoid stale benchmark-only naming drift.

Validation

• Run the relevant benchmark_physics_step_* scripts for at least one overlapping task / batch-size sweep.
• Confirm JSON outputs are produced under benchmark/outputs/physics_step/....
• Confirm the reported metrics are comparable across backends.

Notes

This issue is about benchmark coverage and evidence for support / performance claims, not about changing the env contract or training entrypoints.