Scalable In-Context Learning on Tabular Data via Retrieval-Augmented Large Language Models