GRKV: Global Regression for Training-Free KV Cache Compression in Long-Context LLMs | ArxivCSExplorer