SAC-ViT: Semantic-Aware Clustering Vision Transformer with Early Exit