Hierarchical priors for Function learning

It is common to learn also the parameters of the GaussianRandomField covariance kernel e.g. the lengthscales appearing within the Matern kernel.
Such parameters can often be jointly learnt with the parameter-learning problem, but this would require a suitable framework to sample from such distributions and build the GRF's internally.

No files or tests are named. Start by locating the GaussianRandomField construction and parameter-learning entry points, then trace how covariance-kernel parameters are represented and sampled. Done means providing a framework that can jointly learn those parameters and build the corresponding GRFs internally.

Written by the indexing model from the issue text.