Towards More Faithful Natural Language Explanation Using Multi-Level Contrastive Learning in VQA