Memory issue for attack using many shadow models or large datasets

# Issue
## Problem Description

Lira attack runs out of memory when using many shadow models
1. shadow models are kept in memory in large list
2. evaluated scaled logits are kept in memory
3. maybe also some issues with GPU memory clean-up

Issue (2) also materialize for large data sets (even for few shadow models).

This limits the usability of LeakPro to small toy problems.

## Expected Behavior

LeakPro should work also for attacks
- on complex models with many parameters
- for large datasets
- using many shadow models.

## What Needs to be Done

Consider splitting the evaluation of the attack in smaller batches
- read shadow model from disk one at a time
- read data batches from disk
- write results to disk
- clean up intermediate results and used models

## How Can It Be Tested or Reproduced

Run attacks to test expected behaviour.