Clean up PR 232: Newton Python 3 fix and fatal diagnostics

[tesslerc]

Follow-up to #232.

Credit: this work builds directly on the original fixes and investigation contributed by @alexhegit in PR #232. Thank you for identifying the Python 3 Newton issue and the need for better fatal-crash diagnostics.

This follow-up keeps the substantive fixes while removing leftover debug instrumentation:

• Convert the Newton dict_keys view to a list before passing it to ArticulationView.
• Enable Python's standard-library faulthandler in train_agent.py for native simulator/CUDA crash diagnostics.
• Configure training logs consistently with inference_agent.py.
• Remove debug prints and duplicate imports from the proposed implementation.
• Add a focused regression test for the Newton Python 3 compatibility fix.

Verification:

• Newton uv environment smoke test initialized the G1 simulator and executed the FK path using Warp 1.12.1, Newton, and Torch CUDA 12.8.
• Newton's match_labels rejects the old dict_keys value with TypeError and accepts the list form.
• Focused regression test: 1 passed.
• Train/SLURM tests: 21 passed.
• py_compile: passed.
• git diff --check: passed.

The FK harness reports a pre-existing 0.000012 m tolerance-edge body-position discrepancy, but the Newton process initializes and runs successfully. No new dependencies are introduced.