Quantifying Attention Flow in Transformers