Generalizing support for categorical data

[jbednar]

Right now, we offer the count_cat reduction operator, which effectively does the count reduction separately per category. Most of the other reduction operators would also be meaningful to do categorically, and it would be a pain to implement a categorical version of each one as it is needed.

Instead, we should be able to implement a general-purpose categorical operator, similar to the summary operator, which takes a scalar reduction operator and applies it per category. @jcrist suggests a syntax like by('category_column', mean('some_other_column')), which seems clear to me as well. I think we should do this as soon as we need to apply any operator other than count to categories, since making a general version is probably not much more difficult than copying count_cat and changing the counting bits.