What are the key advantages of Transformer architectures over traditional sequence-based models like RNNs and LSTMs, particularly in terms of efficiency, scalability, and handling long-range dependencies?