feat: Pathwise sampling / Matheron's rule refactoring.

[daniel-dodd]

The proposal here is to extend and structure #195 as a module than just a method on the GP. It is still nice to have quick access to this, but structuring it this way would:

(1) facilitate enhanced testing - there is quite a bit of code inside the function that is not really effectively covered by tests.
(2) facilitate this rule for non stationary kernels and for kernels that are already decomposed as basis functions. Currently we only support this if the Kernel has a spectral density!

Essentially we would give this an appropriate (possibly approximate) prior and then be able to get some samples.