Semantic Token Clustering for Efficient Uncertainty Quantification in Large Language Models — ThinkLLM