Prompt caching makes self-consistency cheap for long-context LLMs