Add support for loading model from splits

[pminev]

Support for this api:
https://github.com/ggml-org/llama.cpp/blob/1a24c4621f0280306b0d53a4fa474fc65d3f1b2e/include/llama.h#L428

Now they have still support for that internally, but since they have implemented the API I suppose they will might remove it. However, it's better to be explicit to the clients and expose the API for them.