Implement message passing for marginalisation of discrete variables

[yebai]

Discrete random variables for PGMs can be marginalised using message passing in exact form. A good example of this is HMMs; see, e.g.,

https://gist.github.com/JasonPekos/82be830e4bf390fd1cc2886a7518aede

Let's implement a general version of the forward-backwards algorithm to marginalise out all discrete variables, enabling us to run HMC/NUTS on the remaining continuous variables.