Implement incomplete cholesky

[karlnapf]

This is another important way to approximate kernel methods

See here for a reference implementation (there are also unit tests)

This book also contains a high level description of the algorithm and matlab pseudo-code

This task is closely related to the random fourier features counterpart, #2768

[karlnapf]

Both this and Nystrom would start with KRR. KRR should put the method to solve the (K+\lambdaI)^-1 system in a method. A subclass would then overload this method and implement the cheaper matrix inversion for Nystrom or incomplete cholesky