Vision Transformers for Dense Prediction