Helpful or Harmful Data? Fine-tuning-free Shapley Attribution for Explaining Language Model Predictions