Support of SMOTE with cross-validation

Hello, congrats for the project. It is purely awesome.

I have a small enhancement to suggest: Sometimes we have to deal with imbalanced datasets and we employ "Synthetic Minority Oversampling Techniques" (SMOTE). However, oversampling needs to only be done in the training set and not in the validation/test sets. Thus, in a cross-validation scenario, at each split SMOTE needs to be done only on the train split and not on validation split.

This functionality is not currently supported by FLAML. I would suggest the SMOTE over-sampler object be passed in the automl_settings and then used at each split.

The imbalanced-learn package achieves the same functionality using custom sklearn-like Pipelines.