Fast approximations

The aim of this project is to implement fast methods for nearest neighbor search and point cloud convolution in a way that is 100% generic and transparent for the wider ML community.