perf: optimize issue 673 reset_done paths

[TATP-233]

Summary

Fixes #673.

• Use get_body_pose_w_rows(env_ids, body_ids) in G1 motion-tracking reset observation so reset only reads the sparse rows being reset instead of materializing full-batch body pose.
• Reuse the Motrix SceneData reset slice when refreshing link-pose cache, avoiding a second self._data[mask] in the reset path.
• Use contiguous qpos slices for Motrix reset actuator ctrl when joint indices are contiguous, with existing fancy-index fallback for non-contiguous layouts.
• Remove the temporary reset attribution/profile scaffolding before PR; final file diff only keeps the owner-layer optimizations and regression test.

Performance Notes

The detailed report was posted to #673: #673 (comment)

Same-process A/B attribution before removing the temporary harness showed:

Path	Backend	Metric	Old	New	Speedup
Body pose reset obs	Motrix	body_pose_ms	1.596 ms	0.064 ms	24.76x
Body pose reset obs	MuJoCo-uni	body_pose_ms	1.446 ms	0.111 ms	13.01x
Motrix set_state	Motrix	set_state_ms	2.320 ms	1.794 ms	1.29x
Motrix set_state	Motrix	refresh_cache_ms	0.638 ms	0.135 ms	4.71x

Overall collector throughput is intentionally not claimed as a PR result because backend_physics_ms drifted between baseline and optimized runs. The supported claim is limited to the local reset paths above.

Validation

• uv run ruff check src/unilab/base/backend/motrix/backend.py src/unilab/envs/motion_tracking/g1/tracking.py tests/envs/test_env_configs.py tests/dr/test_manager.py benchmark/benchmark_offpolicy_collector_active.py
• uv run pytest tests/envs/test_env_configs.py::test_g1_motion_tracking_reset_observation_uses_sparse_body_pose_rows tests/base/test_motrix_backend_options.py tests/dr/test_manager.py tests/benchmark/test_offpolicy_collector_active_benchmark.py -q
• git diff --check
• make test-all