Improving Grapheme-to-Phoneme Conversion through In-Context Knowledge Retrieval with Large Language Models