Benchmarking LLM-Based Synthetic Data Generators for Structural Coherence in behavioral and Human-centered Datasets