perf: skip substep callbacks for invariant actions

[TATP-233]

Summary

• Add a validated ActionTerm.requires_substep_state_feedback declaration, defaulting to False, and aggregate it in ActionManager.
• Register the existing pre-step callback only when an active term needs fresh state between physics substeps.
• Keep Go2WMixedAction on the callback path; invariant actions are applied once and use the existing backend step(ctrl, nsteps=N) path.
• Cover invariant, mixed-feedback, single-substep, empty-manager, rejection, exception, counter, and close-cleanup behavior.

Linked Work

• Closes perf: state-independent MBA action 跳过 per-substep callback (#1259) #1276
• Parent: Roadmap: 优化必选的 Manager-Based runtime collector 吞吐并完成 #1042 性能验收 #1259
• Evidence: Research: 单次 BatchEnvPool dispatch 保留 MBA per-substep control 的可行性 #1262
• Depends on completed perf: MuJoCo backend 接入单次 dispatch per-substep control (#1259 M1b) #1267

Validation

• make check
• uv run pytest tests/managers/test_core_managers.py tests/envs/test_manager_based_rl_env.py tests/envs/locomotion/go2w/test_go2w_manager_based_flat_cfg.py -q (86 passed)
• uv run pytest tests/base/test_backend_pre_step_control.py tests/base/test_mjwarp_capabilities.py -q (13 passed, 1 deselected)
• make test-all (2216 passed, 28 skipped, 275 deselected, 1 xfailed)

Final commit: 27e092bcadb47f0c9379375f21c8dc691b75be95

Benchmark

Fixed command on the same host for both SHAs:

uv run scripts/benchmark/rl/benchmark_offpolicy_collector_active.py --cases flashsac/g1_walk_flat/mujoco --num-envs 8192 --warmup-steps 10 --measure-steps 100 --replay-capacity-steps 64

Metric	Before 70836033	After 27e092bc	Delta
Collector throughput	68,579 env/s	88,683 env/s	+29.3%
Env step	118.420 ms	91.330 ms	-22.9%
Backend physics	99.309 ms	73.716 ms	-25.8%
Update state	8.458 ms	8.400 ms	stable
Reset done	5.172 ms	5.228 ms	stable

The timing shift stays in the expected action/dispatch boundary. Raw local JSON was written to /tmp/issue1276-{before,after}.json and is not committed.

Impact

• Backend impact: MuJoCo invariant-action MBA tasks move from callback dispatch to the existing broadcast trajectory path. Motrix and Drake reuse their existing direct nsteps paths. No backend adapter or SimBackend method changes.
• Go2WMixedAction remains state-feedback driven and keeps per-substep joint-state reads on MuJoCo, Motrix, and Drake.
• mjwarp registration/config/support evidence is unchanged. Its configured-only G1 owner may avoid the callback gate, but this PR does not promote support; local runtime smoke could not proceed because the optional mujoco-warp dependency is not installed.
• Platform impact: Linux validation and benchmark.
• Training effect: collector throughput only. Policy I/O, action values, rewards, reset/RNG, control decimation, and sim2sim fields are unchanged.

Checklist

• Added tests at the action-manager and env callback-selection boundaries
• Linked the driving issue
• Recorded MuJoCo and non-MuJoCo behavior impact
• No task/backend allowlist, new execution path, or support claim added