Try kdtree for large datasets to avoid iterating over all points

[lzkelley]

This should definitely work for finite kernels, and could be a fair approximation even for infinite ones.