The naive_bayes needs some love!

1. The current implementation of BaseDiscreteNB requires that the inputs are positive because otherwise the log computation in line 275 will fail. If BaseDiscreteNB is only applicable to non-negative inputs we should enforce that.
2. GaussianNB what if the variance of a feature is zero? I did a quick fix and add a small constant epsilon (=1e-9).
3. GaussianNB and BaseDiscreteNB should provide similar functionality. E.g. GaussianNB lacks a fit parameter class_prior and class_weight.
4. GaussianNB has a performance regression on bench_covtype - error rate used to be 0.23 now 0.46 - this needs to be investigated.
5. GaussianNB has poor test time performance - should be similar to SGDClassifier (use fortran-style for model parameters).