Consider adding static method to VersionedModel to load best model from a collection of paths

[kostaleonard]

We want to make it easy for users to get their best VersionedModel from their published paths, so we initially included the following static method prototype:

    @staticmethod
    def get_best(paths: list[str],
                 metric: str = 'val_loss',
                 comparator: Callable[[Any, Any], Any] = min) -> \
            'VersionedModel':
        """Returns an instance of the best VersionedModel from the given paths.

        :param paths: A list of the VersionedModel paths to check. Each path may
            be on the local filesystem or remote, independent of the other
            paths. An S3 path should be a URL of the form
            "s3://bucket-name/path/to/dir".
        :param metric: The name of the metric in the model's training history
            to use as the basis of comparison for determining the best model.
            This metric is often the validation loss.
        :param comparator: The function to use to compare model metric values.
            It takes two model metric values and returns the more desirable of
            the two. For example, min will return the model with the lowest
            metric value, say, validation loss.
        """

However, it's not clear whether we have enough information to choose for the user which of their models is the best. Should we pick the model based on the metric score in the last epoch? In the epoch that scores best according to the comparator? Should we just load all of the models and evaluate them on the validation dataset? This could be a helpful feature, but we can't make assumptions on the user's behalf. In the first release, the user can choose for themselves how to compare models.

Also included were the following test function prototypes:

def test_get_best_model_gets_min_val_loss() -> None:
    """Tests that get_best returns the versioned model with the minimum
    validation loss."""
    # TODO
    assert False


def test_get_best_model_custom_comparator() -> None:
    """Tests that get_best returns the versioned model with the maximum
    validation loss when max is supplied as the custom comparator."""
    # TODO
    assert False


def test_get_best_model_custom_metric() -> None:
    """Tests that get_best returns the versioned model with the minimum
    performance on the custom metric when one is supplied."""
    # TODO
    assert False