Added kernel approximation with the Nystroem sub-space projection method as an alternative

Setting the number of nonlinear components, this would allow one to project to a non-linear hyperspace and make a linear fit in it. It has the advantage that the linear fit is still done, but after a nonlinear transformation.

Additionally, an automatic relevance determination linear fit is done instead of the standard Bayesian fit, which should improve resilience to overtraining.