Support split-mode=row for nvlink optimization

### What is the issue? I have read every issue and commit related to Nvlink and I seem to be running in circles. - Is there a way to activate --split-mode=row and get the benefits of NVLink? - What about --tensor_split? - Are there any other options or features that leverage NVLink? _I have found split-mode in this file:_ https://github.com/ollama/ollama/blob/ff7c9060ec62fc3ce5fdc9cbec0788d46d908d81/llm/ext_server/server.cpp#L2345-L2367 ### OS Linux ### GPU Nvidia ### CPU Intel ### Ollama version 0.3.2