A comparative study of zero-shot inference with large language models and supervised modeling in breast cancer pathology classification