Likelihood/Loss classes

[aleximmer]

Probably subclass from torch criteria and keep module parameters for specific library functions. Additionally could subclass from torch distributions for log probabilities and implement the predictive etc.