An Empirical Study of Sample Selection Strategies for Large Language Model Repair