templatize classifier::Kernel

[Peter]

Inspired by NCC, I think we should templatize classifier::Kernel.

There are two choices. First, if the kernel matrix should be calculated on-the-fly or calculated once and cached. Second, what type of KernelFunction is used to calculate an element in kernel matrix.

These choices are currently allowed through dynamic polymorphism. The KernelFunction structure resembles a lot from how Distance previously worked for NCC. For the first choice, typically you want to cache the matrix, and therefore it would be nice to templatize this and make caching default.

I am not sure how far the templates propagate, but I would guess also KerneLookup need to be templatized, and since SVM holds a KernelLookup. This is a bit unclear though, especially since ticket:287 is not finished.

[Peter]

removed CamelCase links