Add Maximum Mean Discrepancy metric

[Warvito]

Another usual metric to evaluate the performance of generative models is the Maximum Mean Discrepancy. Similar to how it is used in this paper https://arxiv.org/abs/2209.03177, it can quantifies the distance between the distributions of real data vs synthetic data.

[danieltudosiu]

@Warvito some design questions:

1. Should we make both a loss and a metric?
2. Should we make the metric/loss kernel agnostic? Similar to how we used it in the NMI submission.
3. Should we make the metric similar to this implementation?