What does this PR do?

This PR adds artifact logging support to the MLFlowLogger by implementing the upload_file() method from the LoggerDestination interface.

Key Changes:

• Implements upload_file(): Adds the ability to log files and directories as MLflow artifacts
• Smart file/directory detection: Automatically uses MlflowClient.log_artifact() for single files and MlflowClient.log_artifacts() for directories
• Distributed training support: Properly respects the _enabled flag to ensure artifacts are only logged from rank-zero in distributed settings
• Implements can_upload_files(): Returns True to advertise the artifact upload capability

Motivation:

Previously, MLFlowLogger could not upload arbitrary files as artifacts (e.g., checkpoints, model weights, configuration files). Using mlflow.log_artifact() directly in training code caused conflicts in distributed training scenarios with world_size > 1. This implementation provides a proper, safe way to log artifacts that works correctly in both single-node and distributed training.

Implementation Details:

The upload_file() method:

• Accepts both file_path pointing to files or directories
• Preserves directory structure when specified in remote_file_name
• Validates that the MLflow run is initialized before logging
• Returns early when _enabled=False (distributed training, non-rank-zero)

What issue(s) does this change relate to?

• Related to artifact logging in distributed training environments
• Addresses the gap where MLFlowLogger couldn't upload files through the standard LoggerDestination interface

Before submitting

• Have you read the contributor guidelines?
• Is this change a documentation change or typo fix? If so, skip the rest of this checklist.
• Was this change discussed/approved in a GitHub issue first? It is much more likely to be merged if so.
• Did you update any related docs and document your change?
• Did you update any related tests and add any new tests related to your change? (see testing)
  • Added test_mlflow_upload_single_file() - tests single file upload
  • Added test_mlflow_upload_directory() - tests directory upload with multiple files
  • Added test_mlflow_upload_file_not_enabled() - tests distributed training behavior
  • Added test_mlflow_can_upload_files() - tests capability advertisement
• Did you run the tests locally to make sure they pass?
• Did you run pre-commit on your change? (see the pre-commit section of prerequisites)

Testing Notes:

All tests follow existing patterns in test_mlflow_logger.py:

• Use pytest.importorskip('mlflow') for conditional test execution
• Use tmp_path fixtures for isolated testing
• Verify artifacts are correctly stored in MLflow directory structure
• Test both success paths and edge cases (disabled logger)