SCNet: Enhancing Few-Shot Semantic Segmentation by Self-Contrastive Background Prototypes