On the Effectiveness of LayerNorm Tuning for Continual Learning in Vision Transformers