DeiT-LT Distillation Strikes Back for Vision Transformer Training on Long-Tailed Datasets