Convolutional neural networks represent shape dimensions—but not as accurately as humans