[WIP]Feat: Run checkpoint ONNX export in a detached subprocess and dedupe save_utd/best exports

[yamsam]

Summary

• Moved the per-checkpoint ONNX export out of the training process into a detached, CPU-only
  subprocess (launch_checkpoint_onnx_export + a new python -m diffusion_planner.utils.onnx_export
  CLI). The export already rebuilds the model from the on-disk .pth + args.json, so it needs
  nothing from the training process — rank 0 no longer stalls all ranks behind the CPU trace of 4
  graphs (the full graph contains 10 unrolled DiT evaluations)
• Deduplicated the double export: when an epoch is both a save_utd snapshot and the new best model,
  the checkpoint is now traced once and the artifacts are copied to the second directory (previously 8
  traced graphs, now 4)
• The child runs with CUDA_VISIBLE_DEVICES="" (never touches GPU memory), logs to
  <output_dir>/onnx_export.log, and is detached via start_new_session; the training loop reaps
  finished exports each launch and waits for stragglers before exiting so artifacts are always complete
• Side effect: the export's global RNG seeding (random/np/torch.manual_seed(42) in
  export_checkpoint_onnx) no longer runs inside the training process, so it can no longer reset rank
  0's augmentation RNG mid-training (O-B1 in docs/onnx_export_analysis.md)
• export_checkpoint_onnx_guarded is kept unchanged for existing callers (e.g.
  train_grpo_predictor.py); migrating GRPO can follow up separately

Test plan

• Added diffusion_planner/tests/test_onnx_export_launch.py (5 tests): command/parser round-trip,
  artifact copying (creates missing dirs, includes external-data sidecars, excludes unrelated files),
  main() exports once then copies, launch is detached and CPU-only with a log file, launch failure
  returns None without raising
• PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 uv run pytest diffusion_planner/tests/test_onnx_export_launch.py
  — 5 passed
• Verified python -m diffusion_planner.utils.onnx_export --help works and diffusion_planner.train
  imports cleanly