Exploring the Robustness of In-Context Learning with Noisy Labels