Exploring Concept Depth: How Large Language Models Acquire Knowledge at Different Layers?