UniPruning: Unifying Local Metric and Global Feedback for Scalable Sparse LLMs