How does the Transformer architecture capture long-range dependencies better than RNNs and LSTMs?