Evaluation metrics can return dict

[guillaumejaume]

Currently a metric must return a float in order to be properly logged to MLflow. I suggest that we had the possibility that a metric returns a dict with float as values. Typically to save confusion matrices, per class accuracy etc...

For a given metric called my_metric, we would log it using the metric name and the keys of the returned dict as my_metric/submetricA.

Also as it is not clearly said that the metrics need to return floats (or dict of floats), we could have a proper error handling.