Important optimization for pan-allele models: pass network the 'name' of the allele rather than the sequence. Network uses a fixed non-trianable allele encoding that maps allele 'names' to sequences.

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

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

This builds on #114