Adaptive Pruning for Large Language Models with Structural Importance Awareness