Augmentation Matters: A Simple-Yet-Effective Approach to Semi-Supervised Semantic Segmentation