Modify Kronecker Gaussian kernel to take advantage of factorized kernel.

[jsjol]

Make use of the special structure in spatial covariance as described in #13

[jsjol]

If you want to use a Kronecker factorization of the spatial covariance, then inputs need to be on a grid.

[jsjol]

My conclusions so far:

• The function gaussian_grid_inference is where most of the action is. Call chain is as follows: model = gp_grid_regression -> gp_grid -> gaussian_grid_inference.

• gaussian_grid_inference is currently hard-coded to use rbf covariance.

• The gradient update poses the main challenge. Because the full matrix K is so big, it is not possible to explictly compute dL_dK and pass that around the way GPy usually does it.

[jsjol]

Need kernels to be able to provide dK_dTheta.

[jsjol]

Code in need of testing:

• stationary:dK_dParams with ARD=True