Add option to disable KV offload

[nsntiw]

Llama.cpp's -nkvo argument disables KV cache offload to the GPU, this allows a 9B model to be loaded onto 9GB of VRAM and used without risk of OOM as context increase. This is very important as R1 distilled models often reason with thousands of tokens before giving an answer.

https://github.com/ggerganov/llama.cpp/blob/master/examples/server/README.md