Experiment with permuting the point cloud data according to KDTree index structure

[dokempf]

This is a rough idea that could be worth exploring to increase the cache locality of our computations: KDTree internally stores a vector of point indices. This index vector could be replaced by the identity vector, if the input point cloud is permuted according to the index vector after the search tree is constructed. This could also be done explicitly from the outside e.g. by something like epoch.optimize_memory_layout().