Avoid duplication of MLflow experiments in array jobs

[sfmig]

When running a training or evaluation array job with a new MLflow experiment name, there is a race between the first few jobs in the array, and a couple of experiments with that name will be created.

I need to look into it, and see if it's possible to avoid this by:

• forcing a pause after the first job?
• creating an experiment if it doesn't exist before launching the training job