Preserving Commonsense Knowledge from Pre-trained Language Models via Causal Inference