fix(appo): keep Motrix playback observations in tensors

[TATP-233]

Reviewable outcome

The APPO/Motrix playback loop now keeps policy I/O on the strict TorchEnv tensor boundary:

• reset indices are allocated on env.device as int64 tensors;
• actor observations stay float32 Torch tensors and fail closed if a non-tensor reaches the reset or step boundary;
• TensorDict policy input is moved explicitly to the actor device;
• actor output is moved directly to authoritative env.device as contiguous float32 and submitted to TorchEnv.step();
• the next actor observation is extracted from the returned TorchEnvState without NumPy or a host roundtrip.

Closes part of #1704.

Validation

• UNILAB_LOCAL_UNISIM=/home/user/ws/unilabsim/unisim uv run pytest tests/scripts/test_train_scripts.py -q -k 'run_motrix_play_loop' --override-ini='markers=' -m '' — 1 passed.
• UNILAB_LOCAL_UNISIM=/home/user/ws/unilabsim/unisim make check — pass (mypy clean; one pre-existing unresolved optional drake_uni.runtime pyright warning).
• UNILAB_LOCAL_UNISIM=/home/user/ws/unilabsim/unisim make test-all — 1834 passed / 26 skipped, benchmark imports 35/35 and 36/36.
• Full slow selection for tests/scripts/test_train_scripts.py has three pre-existing failures on the unchanged base (two off-policy video return assertions and one play_interactive physics-state contract failure); make test-all deselects that slow lane and the focused changed test passes.

Notes for review

Independent of #1725–#1730 and based on current develop/tensor-runtime.