Large language models can be zero-shot anomaly detectors for time series?