Add run-time feature training for ALS

[mdekstrand]

The ALS algorithms are readily amenable to computing new or updated feature vectors for new users on the fly: just solve the least squares problem for the user's feature vector given their ratings and the item feature matrix.

We just need to write the code, and possibly factor out part of the solver to allow single-vector solutions. This is probably most robust with the LU-decomposition solver, although we could run the coordinate descent solver for a few rounds.

Unlike #60, this pertains to what we do at predict or recommend time, not updating the model itself.

[mdekstrand]

We can do this by creating single-row versions of the LU decomposition code.

For example, for BiasMF, we could write a function:

def _train_bias_row_lu(items, ratings, other, reg):
    """
    Args:
        items(np.ndarray[i64]): the item IDs the user has rated
        ratings(np.ndarray): the user's (normalized) ratings for those items
        other(np.ndarray): the item-feature matrix
        reg(float): the regularization term

    Returns:
        np.ndarray: the user-feature vector (equivalent to V in the current LU code) 
    """
    pass

This function probably does not need to be accelerated with Numba.

[mdekstrand]

Re-opening because we still need to do ImplicitMF.