Why neural networks find simple solutions: The many regularizers of geometric complexity