bench(rl): keep active collector transitions in tensors

[TATP-233]

Summary

• Allocate off-policy active-benchmark actions and episode bookkeeping as Torch tensors on env.device.
• Keep observations, rewards, done, truncation, terminal masks, and terminal observations as Torch tensors through replay ingress.
• Remove the benchmark-side NumPy terminal-observation helper and all torch.from_numpy transition conversions.
• Keep host publication only for terminal episode scalar logging.
• Add an AST regression test that rejects NumPy transition carriers and the NumPy terminal-contract helper in the active collector loop.

Part of #1704.

Stacked on #1726 (which is stacked on #1725). Merge in that order, or retarget after the preceding merges.

Validation

• UNILAB_LOCAL_UNISIM=/home/user/ws/unilabsim/unisim uv run pytest -q tests/benchmark/test_offpolicy_collector_active_benchmark.py tests/benchmark/test_extract_offpolicy_metrics.py — 35 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 — 1835 passed, 26 skipped; benchmark imports 35/35 and 36/36

Remaining boundary

ReplayBuffer is currently CPU-backed in this benchmark, matching its existing CPU ingress contract. uni_rl still needs its tensor-first public env contract and single tensor collector path; that external package work remains tracked by #1704.