Differentially Private CutMix for Split Learning with Vision Transformer