Not reusing shadow models

# Issue
## Problem Description
When training shadow models and not using "training_data_fraction" = 0.5, e.g. 0.72, the shadow models is always retrained because they are trained with 0.5.

## Expected Behavior
Should use "training_data_fraction" when selecting amount of data points to use when creating shadow models.

## What Needs to be Done
Make construct_balanced_assignments() compatible with other splits than 50/50

## How Can It Be Tested or Reproduced
Run examples/mia/cifar_main.ipynb and change training_data_fraction to something else than 0.5