tqkit. Unified toolkit for benchmarking and integrating TurboQuant+ KV-cache compression across llama.cpp, vLLM, MLX, and vllm-swift.