Classifiers that respect categoricals

[shaycrk]

This PR creates a new pair of classifiers, CatInATree and CatInAForest, that respect categoricals when randomly sampling features. A few notes:

• To make use of these, it also adds a couple of utilities to detect columns that should be grouped together as categoricals and modifies the _train method of ModelTrainer to pass these detected column groupings.
• Note that because detecting categoricals relies on the feature config, it will fail to properly group all of the categoricals if column names are getting truncated.
• The implementation here is to simply select a random subset of features that ensure all categorical values are selected together and then train a decision tree using the entire pre-selected subset throughout (as well as using this modified decision tree in a bagging classifier).
• It does not, however, subset the features at each node as a random forest would -- this would require a more substantial rewrite of the guts of sklearn's decision trees, but would be a useful future project.

[codecov-io]

Codecov Report

Merging #34 into master will increase coverage by <.01%.
The diff coverage is 90.72%.

@@            Coverage Diff             @@
##           master      #34      +/-   ##
==========================================
+ Coverage   91.23%   91.23%   +<.01%     
==========================================
  Files          13       13              
  Lines         844      993     +149     
==========================================
+ Hits          770      906     +136     
- Misses         74       87      +13

Impacted Files	Coverage Δ
catwalk/utils.py	82.67% <100%> (+5.12%)	⬆️
catwalk/model_trainers.py	93.19% <100%> (+0.99%)	⬆️
catwalk/estimators/transformers.py	86.56% <80%> (-13.44%)	⬇️
catwalk/estimators/classifiers.py	92.92% <92.85%> (-0.27%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 9c08327...8a72cbf. Read the comment docs.

[shaycrk]

@hakoenig made the good point that with the new imputation branch we should also ensure that imputed flags always get picked up along with their underlying columns, so pushed a quick change that adds those groupings in here as well.