Pause-Tuning for Long-Context Comprehension: A Lightweight Approach to LLM Attention Recalibration