Improve train/test split

[mastoffel]

There's a lot to improve in our training / test set split in data_splitting.py:

• the data is always split. However, we would want test_set_size=0 to work, so there's no split and AutoEmulate uses the full data as training data. We then have to decide what to do with evaluate, as there's no test set to evaluate on. We could allow it to take an external test set, otherwise throw an error. Similarly, refit should throw an error, as the models are refitted on the full 'training' data after cross-validation already.
• data is randomly shuffled before being split. For temporal data, spatial data, grouped data etc. we want more sophisticated splitting techniques.