Beyond Size: How Gradients Shape Pruning Decisions in Large Language Models