Do Laplace approx on hyperparameters for increasing posterior variance

[duvenaud]

Using trick from here:

http://www.robots.ox.ac.uk/~mosb/papers/bbq_nips_final.pdf

But with full Hessian instead of just diagonal...