Interpretability Target_Idx / Multilabel setting is not well-defined

We need to standardize how we do interpretability with multilabel setting:

- Decision boundary needs to be explicitly defined
- Target_Idx is irrelevant for binary classification
- defaullt behavior is not well-documented in the interpreter for multilabel.


https://github.com/sunlabuiuc/PyHealth/blob/master/pyhealth/interpret/methods/integrated_gradients.py 