Adding Priors

[mortonjt]

This pull request does the following

• Add priors for match and gap scoring : this will handle some of the identifiability issues, while allowing some match scores to be negative (Add priors for match / gap scores  #66)
• Add multihead attention layer : this is more or less a multilinear layer that imitates a Gaussian mixture model when generating the scoring matrices (Add another attention layer #67)
• Add gap masking : this will significantly reduce the memory burden from specifying which positions in the alignment matrix to mask(Reduce memory footprint #63)

This also fixes #65

[mortonjt]

We're closing this PR, since it looks like a dead end.