Examining Forgetting in Continual Pre-training of Aligned Large Language Models