ParaICL: Towards Robust Parallel In-Context Learning