attention-heads as samples from posterior distribution in a Bayesian sense

https://aclanthology.org/2020.emnlp-main.17.pdf

Though I think CrabNet might need to be refitted for new samples (i.e. if you specify `N=10`, then you only get `10` samples from the posterior, to get more would probably require refitting, and not sure if these would be directly comparable to the `10` from the first run). Also not exactly sure how this could be converted to individual predictions. Maybe just some basic plumbing in and after:

https://github.com/anthony-wang/CrabNet/blob/9e0d79c5bff56ceae0600015942c54214d78152f/crabnet/kingcrab.py#L151-L157