Scaling Vision Transformers to 22 Billion Parameters