The more, the better? Evaluating the role of EEG preprocessing for deep learning applications