Can LLMs Generate Reliable Test Case Generators? A Study on Competition-Level Programming Problems