Why Batch Normalization Damage Federated Learning on Non-IID Data?