Better patching using LLM prompting, via Self-Consistency