Factor similarity matrices out of item-item CF

[mdekstrand]

Right now the item-item CF implementation combines both item-item k-NN, and computing the k-NN matrix from rating matrix rows (collaborative filtering).

There is no reason these have to be tied together. We should support non-CF similarity matrices.

~Refactor the item-item CF implementation so that similarity matrix computation is a separate component, and can be replaced with alternative implementations much more easily. And make the file smaller.~

[mdekstrand]

Abstracting different multiplications behind CSR has helped a lot with this. I am going to update this to focus on the remaining issues.