bench(env): consume public step timing without private probes

[TATP-233]

Summary

• Add per-phase process CPU-time measurements to the public TorchEnv.step() timing schema, paired with existing wall timings.
• Rewrite the phase-CPU benchmark to consume only state.info["timing"] after public env.step() calls.
• Remove benchmark probes of env._backend.step, env._backend._n_threads, and env._reset_done_envs.
• Submit benchmark actions as float32 Torch tensors on env.device.
• Add lifecycle timing coverage and a benchmark AST contract that rejects private attribute access or timing wrappers.

Part of #1704.

Stacked on #1725; merge #1725 first, or retarget this PR after that merge.

Validation

• UNILAB_LOCAL_UNISIM=/home/user/ws/unilabsim/unisim uv run pytest -q tests/base/test_torch_env.py tests/benchmark/test_env_step_phase_cpu_public_contract.py tests/benchmark/test_env_step_config_contract.py — 26 passed
• UNILAB_LOCAL_UNISIM=/home/user/ws/unilabsim/unisim uv run make check
• UNILAB_LOCAL_UNISIM=/home/user/ws/unilabsim/unisim uv run make test-all — 1834 passed, 26 skipped; benchmark imports 35/35 and 36/36

Notes

• The timing schema remains owned by TorchEnv; the benchmark does not monkey-patch owner methods.
• backend_* sub-timings continue to come from the backend result through the existing public owner boundary.