DBR: Divergence-Based Regularization for Debiasing Natural Language Understanding Models