Design interface for missing data in Kronecker models

@tagordon, @ericagol: I have a proposal.

I expect that a common use case would be something like Rubin where you won't ever have multiple bands observed simultaneously, therefore there's a big overhead introduced by building the full matrices and then masking. So:

- For the low rank terms, I think that this could be most efficiently implemented by just allowing the model to have variable kernel amplitudes. I would implement this by allowing some `NxJ` matrix `Alpha` which you would multiply into `U` and `V` and then square, sum along the 2nd axis and then multiply into `a`. I think that this would be equivalent to the low rank Kronecker model with missing data.
- For the dense version, I think that things will be a bit trickier and I'm not sure what the best interface is. I think it would be worth working this through carefully and honestly I think that it might be worth writing a paper. It looks to me like we might be able to come up with a pretty efficient algorithm for this and we'd probably have a lot of users!