LIME/Black box

[buj201]

Related to #47 is interrogating black box predictors (perhaps using LIME). It might be interesting to roll this into coverage of feature importance -- highlighting that globally important features might not be locally important. In fact, the LIME vignette for continuous and categorical features here actually uses the adult data, so this would be an easy extension -- we could easily tack LIME (from this tutorial) onto the catboost with/without one-hot encodings.