Add uncertainties where possible

[MatthewJA]

Predictor.predict should return (predictions, uncertainty) (or confidence, spread, whatever). This matches the way scikit-learn deals with multiple kinds of returned prediction (see sklearn.neighbors.KDTree).

[MatthewJA]

Prerequisite for #80 and #50 and #71.

[MatthewJA]

Having some trouble with this and GPy. The documentation on the predict method says that it returns an array of variances (or a matrix of covariances), but it seems to just be returning nan. I can't find a fix for this.

[chengsoonong]

I am not sure whether you are implementing GPClassifier or GPRegression. Classification is a bit of a beast with GPs, so I suggest to focus only on regression

[MatthewJA]

For the record, I have an in-progress version of this issue locally. Will finish that off when possible.