Decouple Attack-Specific Logic from MINV User Input Handler

# Feature
We want to decouple attack-specific logic (such as PLGMI's GAN training) from user-defined input handlers (eg. celebA_plgmi_handler.py in examples). Currently, the input handler requires users to implement a training function specific to PLGMI, which makes it difficult to generalize to other attacks. Instead, the attack classes themselves should handle the data selection and any attack-specific components, while the user only provides a generic training function.

## Desired Behavior / Functionality

1. The user's input handler should not require knowledge of specific attacks, other than the opportunity to set attack parameters if desired.
2. The train_gan() function in the input handler should be completely attack-agnostic.

## What Needs to Be Done

1. Modify the train_gan() function in input handler: remove any assumptions about specific attack details. Perhaps also change name of this function.
2. Update prepare_attack in attack classes.

## How Can It Be Tested

1. Run existing attack(s) with this change. User input handler should not have any PLG-MI specific train logic.