Add capability to handle categorical covariates

[melhemr]

No description provided.

[AbdulkadirA]

@melhemr This PR allows using categorical variables but it is not clear how it behaves and when it will fail. Can you elaborate a bit on the scope of the PR. Sepecifically,

• In the case of a binary categorical variable; what is the difference to manually coding it 0/1 and treat it as numerical?
• What happens when the categories (count and levels) don't match between learn and apply?
• What if in the data set in the apply function contains only a subset of categories?
• What if the same categories are supplied, but the variable type is e.g. string and not categorical? Will it work correctly? What if the ordering is different?

If these issues are not carefully considered, I don't recommend to merge this PR as it may lead to unexpected behavior, incorrect results, and other issues.

[rpomponio]

@AbdulkadirA 's comments are all very relevant. I'm guessing some internal data structure will need to keep track of the original category levels and the encoding (i.e., the way values were mapped to numbers). Then, the apply stage will need to compare the new data to the original data and reconcile differences. The Sci-Kit Learn method preprocessing.OneHotEncoder could be useful here.

Sorry I'm just getting around to seeing this. I greatly appreciate your work!