How Large Language Models Encode Context Knowledge? A Layer-Wise Probing Study