Instruction-Following Pruning for Large Language Models