Few-Shot Class-Incremental Learning via Training-Free Prototype Calibration