GP-NARX improvements

[RJT1990]

Decided to do more work on this before releasing the beta:

• [x] Fix the log-likelihood term
• [x] Implement iterative GP
• [ ] Implement PITC pre-conditioning
• [ ] Refactor to allow custom covariance matrices
• [ ] Custom tolerance bounds for iterative GP

[RJT1990]

Some changes to the plan. Now going for sparse GP instead of iterative implementation. Steps to finish this work off are to:

• Finish testing of inducing point model with different inference options
• Finish cythonization of new kernel functions
• Reformat standard gaussian process regression
• Make sure everything is well documented !