Performance improvement for scikit-learn training

[pettyalex]

These changes help make training faster, and will enable us to train a lot of models quickly.

1. Introduce xopen to work with compressed files for all input.
2. Filter dosage lines using regex to avoid unnecessary splitting of large lines of text.
3. Limit BLAS concurrency using threadpoolctl. My own testing discovered that default settings result in huge wasted CPU time.
4. Make n_jobs a configurable option. We probably want to stay around 2