Self-Data Distillation for Recovering Quality in Pruned Large Language Models