Improve feature selection process

[zurk]

Context: https://github.com/src-d/style-analyzer/issues/595#issuecomment-466353578

Things we want to try:

• [ ] preselect features by hands
• [ ] add Feature selection to GridSearch
• [ ] feature agglomeration
• [ ] Give a try https://www.featuretools.com/
• [ ] Train feature selection only once per repo (pin selection for the next runs)

[m09]

For the last point I was more thinking about selecting over a set of training repos (for example by taking the top 500 most selected features over all the repos selections) but indeed it might also be good to try this version.