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