Discovering interpretable structure in neural spike trains with negative binomial generalized linear models