Sketch-of-Thought: Efficient LLM Reasoning