fix(training): export ONNX at opset 18 and decouple play artifact failures

[TATP-233]

Summary

• export_policy_onnx default opset 17 → 18. With torch 2.14 the dynamo exporter emits the onnxscript function library (e.g. aten_isnan from torch.nan_to_num in the uni_rl SAC actor wrapper) at opset 18; requesting 17 made torch version-convert the model back while the function stayed at 18, and the optimizer InlinePass crashed with an opset mismatch — killing the play phase right after training finished.
• Added nonfatal_play_step (unilab.training.run) and wrapped both post-training artifacts — ONNX/JIT policy export and video rendering — in it across all three play paths (train_offpolicy.py, train_appo.py, train_rsl_rl.py), so one artifact's failure no longer blocks the other or crashes the play command; failures print a warning plus traceback and the run continues.
• Training behavior unchanged; only the play/export phase is affected.

Linked Work

• Issue: fix(training): ONNX export crashes end-of-training playback; decouple export and video render failures #1654
• Parent roadmap (when applicable):
• Roadmap declared base (when applicable):
• Milestone:
• Base branch: main

Validation

• make test-all passed on the final local head before this PR was created or updated
• Additional task-specific validation listed below

Commands actually run:

make check                       # ruff + mypy + pyright: all pass
make test-all                    # exit 0 (incl. slow suite; benchmark module-mode 35/35, script-mode 36/36)
uv run pytest tests/training/test_onnx_export.py -q          # 6 passed (incl. new nan_to_num regression test)
uv run pytest tests/training/test_training_helpers.py -q -k nonfatal   # 2 passed
uv run pytest tests/scripts/test_train_scripts.py -m slow -q -k "nonfatal or onnx_export_failure or video_render_failure or skip_onnx"  # 3 passed
uv run train --algo sac --task g1_walk_flat --sim mujoco algo.max_iterations=10
# end-to-end: training -> playback -> policy.onnx exported -> ONNX Runtime verified (max_diff 8.94e-08) -> play_video.mp4 recorded

Remote CI route:

• Base main: current-head CI all pass — https://github.com/Motphys/UniLab/actions/runs/36266050382 (ruff-format, ruff-lint, mypy, pyright, test (ubuntu-slim), benchmark-smoke) and https://github.com/Motphys/UniLab/actions/runs/36266050384 (Build Sphinx).

Impact

• Backend impact: none (export/play orchestration only; verified on mujoco)
• Platform impact: both (verified on macOS; change is platform-independent)
• Training effect expected: no (playback-phase artifact handling only)

Artifacts

• W&B:
• benchmark result:
• video / screenshot: logs/fast_sac/G1WalkFlat/2026-09-27_03-24-29_mujoco/play_video.mp4 (local)
• ONNX / checkpoint: logs/fast_sac/G1WalkFlat/2026-09-27_03-24-29_mujoco/policy.onnx (opset 18, ORT-verified vs PyTorch, max_diff 8.94e-08)

Checklist

• Added or updated tests where needed
• Updated docs if behavior or workflow changed (no doc contract change; opset is not documented anywhere)
• Linked the driving issue
• Noted any follow-up work explicitly