[FEA] Support Naive Bayes variants

[cjnolet]

There are 4 different variants of Naive Bayes in Scikit-learn:

• Multinomial
• Gaussian
• Bernoulli
• Complement

Between experimentation with CuPy RawKernel and abstracting it for type agnosticism, creating a new directory for CuPy/Python-based prims, and initial multinomial Naive Bayes implementation, the Naive Bayes PR (#1375) has become quite large and needs to be merged.

The primary primitive in the multinomial Naive Bayes variant, a custom RawKernel that uses shared memory and atomicAdd to count features for each class, also supports squaring the sums so that it can be used to extract a mean and variance for the Gaussian variant. The remaining variants of the algorithm should also be able to make use of this primitive.

Given the infrastructure provided by #1375, adding these variants should be straightforward and moderate to trivial. The distributed variants should just be able to proxy to the single-GPU classes, combining the underlying parameters in the same way as the existing multinomial version.