Generate and persist UMA model IDs

[lbluque]

Summary

• generate random IDs for untagged UMA MoE models
• broadcast generated IDs from rank zero across distributed workers
• persist resolved IDs in resume, DCP, and inference checkpoint configs
• add IDs to UMA benchmark configs
• add coverage for generation, broadcasting, and checkpoint persistence

Testing

• 33 focused unit tests passed
• DCP train/save/load integration test passed
• pre-commit passed

[wood-b]

look good, I just need a bit more time to think about the changes to compat.py

[wood-b]

do we need to add the model here? I don't love hard coding a model. Also if we do, it should be UMA-S-1.2.1

[lbluque]

That is just setting a model_id for the purpose of testing training. It could be whatever. I changed it to "UMA-s-test" to be descriptive

[github-actions]

This PR has been marked as stale because it has been open for 30 days with no activity.