ProvFL: Client-Driven Interpretability of Global Model Predictions in Federated Learning