Refactoring/Redesign scalar / matrix valued kernels

[marcelluethi]

The base-classes for scalar/matrix valued kernels have currently a few limitations:

1. The base classes are parametrized by the PointType only. For generic implementations of kernels (i.e. those where the concrete type of PointType is not known), the actual dataset type over which the kernel is applied should be known, such that operations onpoints can be performed using the representer.
2. The Kernel combinators are only implemented for MatrixValued kernels. It would be handy to have them also for ScalarValuedKernels
3. The Kernel combinators are difficult to use since they require explicit pointers. Maybe the use of smart pointers would allow for a more flexible and user friendly way to combine kernels