Fix bug in DeepLearning's varimp that happens when max_category_features is in use

Deeplearning with max_categorical_features lower that the number of categorical features fails on varimp.

This can be easily replicated by setting `max_categorical_features=10_000` and running the following dataset: https://www.openml.org/d/41147 .

Possible workaround is disabling the `variable_importances`.