Skeleton-of-Thought: Large Language Models Can Do Parallel Decoding