Exploring Information Processing in Large Language Models: Insights from Information Bottleneck Theory