Text this: Collapse-Aware Clipping for Low-Bit Quantization of Post-Softmax Activations in Vision Transformers