mlp-mixer-vision fully connected architecture (NO convolution) for computer vision that achieves 89% accuracy on CIFAR 10