Max Pooling with Vision Transformers reconciles class and shape in weakly supervised semantic segmentation