Improving LLM Group Fairness on Tabular Data via In-Context Learning