Mitigating social biases of pre-trained language models via contrastive self-debiasing with double data augmentation