Add check for strictly positive training data for CategoricalHMM

[abuzarmahmood]

Apologies if I missed this but I wasn't able to find in the documentation the CategoricalHMM takes positive values only. I only realized it when I noticed my predictions from fitted models behaving weirdly.
It would be useful to add a warning (or even a transform of the inputs) when fitting the HMM.
Happy to submit a PR if this would be useful.