perf: add CUDA graph capture to mjwarp backend

[TATP-233]

Summary

• 在 MjwarpBackend 冷路径检查 CUDA driver >= 12.4 与 Warp mempool，并原子捕获固定地址的 step、forward、reset_data CUDA graphs。
• 稳态 step 按现有 nsteps 重放 graph；selected-row reset 仍严格保持 reset_data -> qpos/qvel upload -> forward -> sync -> host cache 顺序。
• 条件不满足或任一 capture 抛错时，带具体原因发出 RuntimeWarning 并整体回退到原 eager 路径，不留下部分启用状态。
• 不修改 SimBackend、env、runner、learner、配置或其他 backend；本 PR 是 Roadmap: 借鉴 mjlab 的性能机制与开发体验优势（7 个候选 sub-issue） #1252 任务1的 implementation child Perf: mjwarp 后端引入 CUDA graph capture #1271。

Contract / architecture

• 仅改变 mjwarp backend 内部执行机制，SimBackend 的输入、输出、timing keys 与 reset/step lifecycle 不变。
• 现有 hot path 只通过 Warp assign 更新已分配数组，不替换 graph 捕获的地址；代码明确要求未来若 owner layer 替换 model/data array，必须重新调用冷路径 capture。
• MuJoCo、Motrix、Drake、macOS 与非 mjwarp entrypoint 行为不变；不新增 support claim、常规 CI 或长期 benchmark infrastructure。
• 该实现保持 backend capability boundary，不产生新的公共 contract 或 ADR 决策。

Throughput

同机、同任务、同参数分别测量 base 60c2ce7c 与本 PR 653cd7a1：

• GPU: NVIDIA GeForce RTX 4090, 47 GiB, NVIDIA driver 595.84
• Runtime: CUDA Toolkit 12.9；Warp driver API 13.2；warp-lang==1.14.0；mujoco-warp==3.10.0.3；mempool enabled
• Workload: production MjwarpBackend, G1 scene_flat.xml, 2048 worlds, sim_dt=0.02/3, 3 physics substeps/control step, nconmax=128, njmax=256
• Measurement: fresh backend per revision, stand state/control, 10 warm-up calls + 40 measured calls；throughput uses mean full backend.step() wall time and includes control H2D、physics sync、qpos/qvel/sensor D2H cache refresh

Metric	Before	After	Difference
Mean full backend step	10.168 ms	4.582 ms	-54.9%
Median full backend step	9.706 ms	4.287 ms	-55.8%
Mean physics section	9.944 ms	4.392 ms	-55.8%
Control-env-steps/s	201,415	446,996	2.22x / +121.9%
Physics-env-steps/s	604,244	1,340,987	2.22x / +121.9%

The benchmark intentionally reports the production backend barrier rather than the benchmark-only mjwarp adapter. Each control step advances 2048 * 3 physics world-steps.

Validation

• make test-all
  • ruff, mypy, pyright passed
  • 1776 passed, 24 skipped, 273 deselected, 1 xfailed
  • benchmark smoke: module mode 32/33 passed (1 platform-optional skip), script mode 33/34 passed (1 platform-optional skip)
• uv run pytest -q tests/base/test_mjwarp_cuda_graph.py tests/base/test_mjwarp_identity.py — 12 passed
• uv run --extra mjwarp pytest -q -m slow tests/base/test_mjwarp_backend.py tests/base/test_mjwarp_capabilities.py tests/base/test_mjwarp_differential.py tests/base/test_backend_conformance.py — 12 passed, 6 deselected

Closes #1271
Related to #1252