Multivariate Categorical distribution

[mohamed82008]

Currently, I am fairly convinced that the main thing keeping Turing from reaching Stan speeds on https://github.com/TuringLang/TuringExamples/tree/master/benchmarks/hmm_semisup is a multivariate categorical distribution that takes a matrix of probabilities in. Even with ReverseDiff, caching the tape and some broadcasting and vectorization tricks, we are still significantly slower than Stan because of this issue. This issue shouldn't be too hard for anyone willing to start contributing to DistributionsAD.