Boosting Semi-Supervised Medical Image Segmentation via Masked Image Consistency and Discrepancy Learning