Learning Dynamics in Continual Pre-Training for Large Language Models