Data preparation improvements

[miquelramirez]

There are a number of improvements on the data preparation process that will help a lot down the road:

• [ ] Add labels to columns in dataset
• [ ] Refactor prepare_data.py so we use pandas for generating the datasets rather than manually via numpy, which is easier to maintain and to read.
• [ ] Store datasets in compressed format (e.g. gz)
• [ ] Add development option to generate a small sample of the working dataset