Quantizing the KV cache in Ollama on RTX Spark: half the memory, a third slower