Support marginalization of HMM with higher lag orders

[ricardoV94]

Not a blocker, but we should be able to support this. Here's a paper with equations for a 2nd order HMM, we should be able to work out the general formula?

https://journals.sagepub.com/doi/pdf/10.1177/1550147718772541

Originally posted by @jessegrabowski in #257 (comment)

See also comments on unclear API for multiple orders vs batched transition matrices for a given order:
#257 (comment)
#257 (comment)