Rethink the way we are training the shadow models

# Feature
## Desired Behavior / Functionality
While we have changed the training from completely random sampling data point to forcing a dataset split of 50/50 to two different models, it is still not the general case solution. By adopting dataset splitting method like a K-fold style split, making sure that for each data point of the dataset has at least one IN and one OUT model per set of splits, we can support splits that have different fractions like 75/25, 80/20 etc.

We should still support the randomly assigned splits as we did before, and with a solution as mentioned above, that will be an easy implementation.

## What Needs to Be Done