bench(env): use tensor actions for env rollouts

[TATP-233]

Summary

• Allocate benchmark_env_step warmup and measured actions as contiguous float32 Torch tensors on env.device.
• Replace the SimToolReal fixed-tool benchmark NumPy action carrier with a device-resident Torch tensor.
• Validate SimToolReal rollout finiteness with Torch while retaining one explicit scalar publication to the benchmark result.
• Add a focused contract test preventing the env-step benchmark from regressing to np.random.uniform actions.

Part of #1704.

Validation

• UNILAB_LOCAL_UNISIM=/home/user/ws/unilabsim/unisim uv run pytest -q tests/benchmark/test_env_step_config_contract.py — 7 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 — 1833 passed, 26 skipped; benchmark imports 35/35 and 36/36

Remaining #1704 benchmark work

The phase-CPU benchmark private-API probe, NpEnv shard benchmark replacement, off-policy active benchmark tensor transition, and neutral timing metric rename remain open.