Support incremental updates to ALS

[mdekstrand]

Incrementally updating ALS is, in principle, easy:

1. For a new item, solve for the item feature row.
2. For a new user, solve for the user feature row.
3. For new ratings, run a small number of rounds of the solver (maybe just one).

(3) needs a little work - do we need to update all users/items, or just affected ones? We need to store the old ratings matrix, not just the preference matrices, because we need the old data too.

I would propose not updating biases, but letting them get stale.

[mdekstrand]

It is unlikely that we will add actual incremental model update support, but runtime incorporation of new ratings (#114) is definitely in-scope.

Realistic scenarios will usually need concurrency, and model updates are very difficult to do with concurrency. We can re-open if we decide this is useful.