[tool] fix: attach to existing MLflow run when MLFLOW_RUN_ID is set

[dubin555]

Some cloud providers (e.g. Azure ML, Databricks) automatically set the MLFLOW_RUN_ID environment variable.
Previously, verl always called mlflow.set_experiment(...); mlflow.start_run(experiment_id=..., run_name=...), which can conflict with the provider-managed run context and cause MLflow to throw an exception.

This change detects MLFLOW_RUN_ID and attaches to the existing run instead of creating a new one, preventing duplicate runs and enabling seamless integration with managed MLflow environments.

What does this PR do?

• Detect MLFLOW_RUN_ID during Tracking(..., default_backend="mlflow") initialization.
• If MLFLOW_RUN_ID is present, call mlflow.start_run(run_id=...) to attach to the provider-managed run.
• Otherwise, keep the previous behavior: mlflow.set_experiment(project_name) and start a new run under that experiment.

Why is this needed?

On Azure ML, MLflow run context is automatically created and MLFLOW_RUN_ID is injected.
Starting a new run under a different experiment ID triggers a mismatch error like:

MlflowException: Cannot start run with ID <...> experiment ID does not match environment run ID.
Make sure --experiment-name or --experiment-id matches experiment set with set_experiment()...

Because the cloud provider-generated MLFLOW_RUN_ID is not predictable and is set outside the user config, this cannot be reliably worked around via configuration alone.

Test

• ✅ Local smoke test (manual): verified the MLflow backend initializes correctly when:
  • MLFLOW_RUN_ID is set → attaches to existing run
  • MLFLOW_RUN_ID is not set → creates/uses experiment and starts a new run
• CI: (existing unit tests)

(If you have an Azure ML job run link/log snippet, you can paste it here as additional evidence.)

API and Usage Example

No API changes. Behavior is automatically enabled when MLFLOW_RUN_ID is set by the environment.

Design & Code Changes

• verl/utils/tracking.py
  • Add a conditional path to honor MLFLOW_RUN_ID and attach to an existing MLflow run.
  • Preserve the previous experiment-based run creation when MLFLOW_RUN_ID is absent.

[gemini-code-assist]

Code Review

This pull request introduces a fix to allow verl to attach to an existing MLflow run when the MLFLOW_RUN_ID environment variable is set, a common scenario in managed ML environments like Azure ML or Databricks. The change correctly detects the presence of MLFLOW_RUN_ID and uses mlflow.start_run(run_id=...) to attach to the existing run, which avoids conflicts from the previous behavior of always creating a new run. The implementation is clean, directly addresses the described issue, and I found no high or critical severity issues in my review.