Fix bug that led to NaN mhcflurry_prediction_high and mhcflurry_prediction_low uncertainty estimates in some cases

Performance enhancement for faster start time: reuse compiled networks even if their regularizations differ. This is safe because regularization should not affect predictions

Docs improvememnts, more flexible Keras dependency, drop typechecks dependency, remove some unused functions

* First cut of supporting quantiles ("percent ranks") per allele. Closes #87.
* Some refactoring of amino_acid.py

This builds on #114