Towards Reliable Medical Unsupervised Anomaly Detection: A Benchmark and Dataset for PET/CT with Cross-Modality Knowledge Distillation Network