[Enhancement] Automatically remove models not feasible to train when high-cardinality classification problem is presented

[gradientsky]

As a user, I would like AutoGluon to automatically remove models, which are not likely to be trained when high-cardinality classification is presented (i.e. 1M rows, 500 classes are likely make CatBoost model not feasible). User should be able to override defaults if the models are mentioned explicitly.