On the Distribution, Sparsity, and Inference-time Quantization of Attention Values in Transformers