Use slices in calculating composed basis gradients rather than zero padding?

[dsteinberg]

We currently zero-pad dPhi in the BasisCat object for updating hyperparameters only depending on one basis. Would it be more efficient to slice instead? Maybe look into this with #106.