Researchers from MIT, NVIDIA, and Zhejiang University Propose TriAttention: A KV Cache Compression Method That Matches Full Attention at 2.5× Higher Throughput