[CLS] Attention is All You Need for Training-Free Visual Token Pruning: Make VLM Inference Faster