Towards Robust Visual Question Answering: Making the Most of Biased Samples via Contrastive Learning