RAT: Bridging RNN Efficiency and Attention Accuracy via Chunk-based Sequence Modeling | Read Paper on Bytez