refactor(motrix): 移除临时 body-state kernel (#1311)

[TATP-233]

Summary

• 删除 perf(body-state): 接通统一 copy contract 并同步优化 MuJoCo/MJWarp #1307 留下的 Motrix 专用临时 Numba body-state kernel、import、TODO 与专用 kernel 测试。
• 恢复 Motrix 0.8.2 既有的 cache-based NumPy copy_body_state_w() adapter，并验证 body order、xyzw→wxyz、caller-owned outputs 与 scratch 复用。
• 不修改统一 SimBackend/Entity contract；MuJoCo/MJWarp 继续使用共享 parallel Numba kernel。
• 不升级 MotrixSim，不对 Motrix 做本轮 benchmark 或性能/support 声明。训练数值语义不变。

Linked Work

• Closes refactor(motrix): 移除临时 body-state kernel 并退出本轮加速范围 #1311
• Roadmap: roadmap: motion tracking Numba 热点与统一 body-state copy 收敛 #1304
• Replaces the de-scoped refactor(motrix): 接入 fused body-state 接口并删除临时 kernel #1308

Validation

• make check
• focused backend/body-state tests
• make test-all
• MuJoCo/MJWarp final roadmap benchmark

Commands actually run:

uv run --no-sync pytest tests/base/test_body_state_copy.py tests/base/test_motrix_backend_options.py tests/base/test_entity_facade.py tests/base/test_mjwarp_backend.py tests/base/test_sim_backend.py -q
make check
make test-all
uv run --no-sync scripts/benchmark/rl/benchmark_offpolicy_collector_active.py --cases sac/g1_motion_tracking/mujoco --num-envs 8192 --warmup-steps 10 --measure-steps 100
uv sync --extra mjwarp
uv run --no-sync scripts/benchmark/rl/benchmark_offpolicy_collector_active.py --cases sac/g1_motion_tracking/mjwarp --num-envs 8192 --warmup-steps 10 --measure-steps 100

Results:

• focused: 66 passed, 229 deselected
• make test-all: 2300 passed, 28 skipped, 281 deselected, 1 xfailed; ruff/mypy/pyright and benchmark import smoke passed
• working tree clean at tested commit a2e08c61bccffb524d071b9947626eda295a62f1
• Remote CI is intentionally skipped and not awaited under the maintainer/user override recorded in roadmap: motion tracking Numba 热点与统一 body-state copy 收敛 #1304.

Impact

• Backend impact: Motrix cleanup only in this child; MuJoCo/MJWarp accelerated paths unchanged
• Platform impact: Linux tested
• Training effect expected: no numerical/contract change; Motrix returns to the correct non-accelerated adapter
• Dependency impact: none; motrixsim-core==0.8.2 pin and lock remain unchanged

Artifacts

Final roadmap benchmark on AMD Ryzen 9 9950X3D2 + RTX 4090, 8192 envs, warmup 10 / measure 100, three independent runs:

Backend	update_state runs (ms)	Median (ms)	env/s runs	Median env/s	vs initial update_state	vs initial env/s
MuJoCo	19.963 / 19.986 / 20.564	19.986	79,053 / 76,895 / 74,282	76,895	29.766 → 19.986 (-32.9%)	70,297 → 76,895 (+9.4%)
MJWarp	18.748 / 18.894 / 19.289	18.894	150,488 / 149,283 / 146,083	149,283	28.704 → 18.894 (-34.2%)	122,524 → 149,283 (+21.8%)

Motrix intentionally has no final performance result under the amended scope.

Checklist

• Added/updated tests at the backend boundary
• Linked the driving issue
• Explicitly recorded the Motrix follow-up de-scope
• No public contract, reward/config, runner/lifecycle, or CI changes