Refactor metrics logging

[Galileo-Galilei]

Currently, the recommended way to log metrics in kedro_mlflow is the MlflowMetricsDataset class.

However:

• this is not entirely tested and some edge cases fell through the code review. We should make test coverage 100% again.
• some end users reported their need to have a MlflowMetricsDataset to log an int/ a list directly within the catalog. I suggest that we move almost all the methods (the ones which interacts with the mlflow backend store) to a mother class MlflowAbstractMetricsDataset, and then make MlflowMetricsDataset and MlflowMetricDataset inherits from this mother class and impement just the specific logging.
• Contrarily to what we choose, it seems that end users prefer having a load_args parameter to specify if they load the whole history (dict of {step: value}), the values (list of int), the last (last int) or the best value.

@akruszewski do you have time/will to address this one as a complement of your previous PR? If yes, add it to the 0.4.0 milestone and assign yourself (even only completing the tests would be really great!).