Efficient Continual Pre-training for Building Domain Specific Large Language Models