New KV cache compaction technique cuts LLM memory 50x without accuracy loss