Restructure kernels to allow vectorization of kernel functions

[costashatz]

We should at some point change the API to allow vectorization of kernel functions. For example, we would like to do: _kernel = K(X,X) where X is a vector of samples and K is the kernel function. Similarly, _dK_dθ = dK(X,X). That way we will be avoiding the for loops and use Eigen's vectorized functions. In some simple benchmarking that I did, this is the main factor of slowing down our code (especially when optimizing the hyper-parameters of the kernel). The reason we cannot do it in limbo right now is that the kernels provide the following functionality: kernel(x1,x2) and gradient(x1,x2) whereas they should provide kernel(X1,X2) and gradient(X1,X2). We could also provide both functionalities; it is certainly doable with Eigen.