perf(env): remove redundant state finite reduction

[TATP-233]

Summary

• Profiling the whole env.step() on the gait owner showed several dominant device-synchronizing roots from TorchEnv._validate_state: aten::is_nonzero, aten::all, and aten::isfinite.
• perf(managers): batch reset and state validation boundaries #1787 tried to batch those checks by stacking per-output isfinite(...).all() reductions, then reducing the stacked scalars. That creates one reduction and host conversion per output plus a second stacked reduction and conversion, exactly the multi-sync pattern it intended to remove.
• Restore the straightforward per-output isfinite(...).all() check. Each output still has one guarded boundary, but the extra stack/reduction and host conversion are gone.

Validation

• UNILAB_LOCAL_UNISIM=../unisim uv run pytest -q — 1844 passed, 26 skipped, 624 deselected
• UNILAB_LOCAL_UNISIM=../unisim uv run pytest -q -m slow tests/base/test_mjwarp_backend.py::test_g1_walk_flat_owner_one_step tests/envs/test_env_configs.py::test_flashsac_g1_motion_mjwarp_tensor_anchor_observations_roll_out — 3 passed
• uv run ruff check ., uv run ruff format --check ., uv run mypy src/unilab, git diff --check — pass

Part of #1811.