Combining Large Language Models and Gradient-Free Optimization for Automatic Control Policy Synthesis