Ability to make child models of ensemble use part of features

## Description
A clear and concise description of what the feature is.

If a ensemble model consists of automm, cat, gbm, and the train dataset have text columns,
make automm use the text features only, except for the other non deep learning model.

I think this is maybe specific to my dataset, whose text feature has too many n grams (256).

I used this text feature but overfitting was detected in tabularpredictor ensemble model.
I think the number of features was too high due to ngrams and so overfitting occured.

So, **It would be good that I can feed text features only to deep learning model(which can handle the text feature without ngram), not to tabular model(catboost, etc)**

- Please indicate which module (`multimodal`, `tabular`, `timeseries`) this proposal refers to.
Tabular.
- If the proposal is about an API modification, provide mock examples, if possible.

Thanks

Start by tracing the TabularPredictor ensemble and its AutoMM, CatBoost, and GBM child models; the issue provides no file or test entry point. Done means allowing text features to be supplied only to the deep-learning child model while other ensemble models exclude them, with coverage for the reported n-gram overfitting scenario.

Written by the indexing model from the issue text.