Learning to Reduce Dual-level Discrepancy for Infrared-Visible Person Re-identification