Less is More: Unsupervised Mask-guided Annotated CT Image Synthesis with Minimum Manual Segmentations