On the Origins of the Block Structure Phenomenon in Neural Network Representations