Extended Cross-Modality United Learning for Unsupervised Visible-Infrared Person Re-identification