Move MLflow import to function scope to avoid import-time dependencies

[Copilot]

Addresses feedback from #440 to move the upload_artifacts_to_mlflow import from module level to function scope, avoiding potential circular import issues and making the dependency explicit.

Changes

• Move from .mlflow_artifacts import upload_artifacts_to_mlflow from module-level (line 11) to inside upload_mlflow_artifacts() function where it's actually used

This makes the import conditional on the function being called rather than when the module is first imported, reducing import-time coupling.

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.

[gphuang]

Yes, it's a valid pattern, but whether it's preferable depends on context:
Pros of moving import inside function (lazy import):
Avoids circular imports - If mlflow_artifacts.py ever imports from global_vars.py, module-level import would fail
Faster module load - mlflow_artifacts.py only loads when upload_mlflow_artifacts() is actually called
Conditional loading - Since MLflow upload only runs at end of training (and only if enabled), the import is truly deferred
Cons:
Against PEP 8 - Python style guide recommends imports at top of file
Slight runtime overhead - Import lookup happens each call (though Python caches, so negligible after first call)
Less visible dependencies - Harder to see module dependencies at a glance
Delayed error detection - Import errors surface at function call time, not module load time