Investigating Cost-Efficiency of LLM-Generated Training Data for Conversational Semantic Frame Analysis