Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification