DTQAtten: Leveraging Dynamic Token-based Quantization for Efficient Attention Architecture