How does prompt engineering affect ChatGPT performance on unsupervised entity resolution?