Beyond Function Calling: Benchmarking Tool-Using Agents under Tool-Environment Unreliability