Handle discrete types better.

[jsalvatier]

I've run into some issues recently with values being of the wrong dtype. I've seen floats show up in traces for discrete variables (which messes up plotting). I've also had a Poisson missing variable treated as effectively continuous.

This was because pandas.DataFrame({'x' : [1,2,3. np.nan]}) is a float column and passing a float column and Poisson doesn't seem to have enough checks to prevent the variable from actually just sampling as a continuous variable.

Also trace backends always seem to create float arrays for storage.

Also Metropolis seems to always return floats.

These bugs combine in annoying ways.

[jsalvatier]

Also related: #655

[jsalvatier]

I think these issues are mostly fixed.