Attribution-guided Pruning for Compression, Circuit Discovery, and Targeted Correction in LLMs