Support for sparse matrices

[slinderman]

This is a rather vague issue... I'm working on a few projects that have very large and very sparse observations. That motivated the development of the ZeroInflatedCountModels, which takes in sparse matrices in csr_matrix format. We could extend this interface to all of the _LDSStates objects, and it might just work™ since we the data are primarily used to compute node potentials, which are mostly just dot products.

Count data present a more challenging problem. If the data are not "zero inflated," I don't see any way around instantiating a matrix of auxiliary variables the same size as the observations.