feat: CUDA port of TurboQuant3 KV cache — 3.47x compression, 98.5% of F16 decode speed on RTX 5090#3

signalnine wants to merge 4 commits intoTheTom:feature/turboquant-kv-cacheTheTom/llama-cpp-turboquant:feature/turboquant-kv-cachefrom

signalnine:feature/turboquant-kv-cachesignalnine/llama-cpp-turboquant:feature/turboquant-kv-cacheCopy head branch name to clipboard