Multicalibration for Confidence Scoring in LLMs