perf(motion): 固定五个 update_state 热点改为 Numba-only

[TATP-233]

Summary

• Replace termination/anchor_pos and the four motion_body_* rewards with pre-warmed njit(nogil=True, parallel=True) kernels.
• Keep the four reward formulas Numba-only in the production call graph; NumPy expressions remain only as test oracles.
• Reuse caller-owned result buffers and preserve body subset, dtype, input immutability, and manager class-term contracts.
• Add Numba as a runtime dependency.

Linked Work

• Resolves perf(motion): 固定五个 update_state 热点改为 Numba-only #1306
• Parent roadmap: roadmap: motion tracking Numba 热点与统一 body-state copy 收敛 #1304
• Follow-up: perf(body-state): 接通统一 copy contract 并同步优化 MuJoCo/MJWarp #1307

Validation

• make check
• uv run pytest -m "not slow"
• make test-all
• Three-backend 8192-env benchmark, warmup 10 / measure 100, 3 runs per side, median of run means

Commands actually run:

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 run --no-sync scripts/benchmark/rl/benchmark_offpolicy_collector_active.py --cases sac/g1_motion_tracking/motrixsim --num-envs 8192 --warmup-steps 10 --measure-steps 100
uv run --no-sync --extra mjwarp scripts/benchmark/rl/benchmark_offpolicy_collector_active.py --cases sac/g1_motion_tracking/mjwarp --num-envs 8192 --warmup-steps 10 --measure-steps 100

make test-all result: 2294 passed, 27 skipped, 281 deselected, 1 xfailed; ruff, mypy, pyright, and benchmark smoke passed.

Maintainer override: local make test-all only; remote CI is intentionally not awaited or required for this roadmap.

Benchmark

Same host: Ryzen 9 9950X3D, RTX 4090. Values are before → after.

Backend	update_state ms	reset_done ms	env/s
MuJoCo	29.766 → 25.778 (-13.4%)	7.031 → 7.077 (+0.6%)	70,297 → 73,628 (+4.7%)
Motrix	38.648 → 35.782 (-7.4%)	12.942 → 13.837 (+6.9%)	65,423 → 66,543 (+1.7%)
MJWarp	28.704 → 25.323 (-11.8%)	14.916 → 14.971 (+0.4%)	122,524 → 130,895 (+6.8%)

reset_done is not modified by this child. MuJoCo/MJWarp are flat; the Motrix delta tracks reset-count/backend run variance and is reported without claiming a reset improvement.

Impact

• Backend impact: mujoco / motrix / mjwarp
• Platform impact: Linux CPU hot path; Numba thread env overrides are respected
• Training effect expected: same reward/termination semantics with lower update_state CPU cost
• Configuration and runner/lifecycle behavior: unchanged

Artifacts

• benchmark result: local gitignored JSON/CSV, 18 runs total
• W&B/video/ONNX/checkpoint: not applicable

Checklist

• Added numerical parity, subset, output reuse, input immutability, and compilation tests
• Linked the driving issue
• Follow-up body-state copy work is isolated in perf(body-state): 接通统一 copy contract 并同步优化 MuJoCo/MJWarp #1307