Feature/gpu priors

[nbarlowATI]

Refactoring the C++/GPU implementation to hold hyperparameters in a GPParams object, and introduce priors for hyperparameters following the same logic as the Python/CPU implementation.

• New C++ classes and corresponding Python bindings:
  • GPParams - holds values of parameters for mean function and kernel parameters (including nugget)
  • GPPriors - container class for Prior objects, including:
  • MeanPriors - not yet used in analytical meanfunc calculation, but still containing much of the necessary functionality.
  • WeakPrior - base class for other prior distributions
  • NormalPrior
  • LogNormalPrior
  • InvGammaPrior
• Default behaviour and interface of GPPriors in the C++ / GPU implementation is identical to that in the Python/CPU implementation: if no GPPriors object is passed to the GP's constructor, default will be InvGammaPrior with appropriately chosen shape and scale for the correlation length parameters
• If a python GPPriors object is provided, python code in GaussianProcessGPU will ascertain what prior distributions and parameters need to be passed to the C++ class for it to construct the corresponding C++ versions.
• The code in fitting.hpp will use the GPPriors::sample method to obtain starting values for the fit, rather than uniform random values.

[edaub]

Mostly looks good -- a few things in the code I though might be cleaner, and a few random thoughts/questions.

Two things I that do need to be fixed:

• The correlation length priors appear to all enforce the same distribution (or I missed something) in GaussianProcessGPU.py

• Clarify what is going on with the distributions in MeanPriors in the C++ implementation (and that currently it only implements weak priors). In particular, I'd get rid of the other distributions and just make it clear with some comments that you need to make a bunch of weak prior objects for sampling purposes.