Uncertainty Decomposition and Quantification for In-Context Learning of Large Language Models