Synthetic Data Generation with Large Language Models for Text Classification: Potential and Limitations