Code style improvements

[helloxinw]

• Move util files from data to a new util folder. Only the clean_* and build_features files should be in data.
• Merge clean_labels and clean_dataset
• Re-organize train.py and predict.py. E.g. separate "evaluate" functions from predict.py
• Define the data types in the labels matrix more explicitly. Converting labels to and from pickle files makes the data type inconsistent.
• Fix inconsistencies in low-pass filters between training and predicting