Performance-Guided LLM Knowledge Distillation for Efficient Text Classification at Scale