Replace hardcoded tweaks to llama.cpp with a more permanent solution

[nsarrazin]

Currently the compiled llama.cpp binary we use only supports alpaca. The source had to be modified to accept a model as a single file (alpaca 13B is a single file, as opposed to the 2-part model expected for LLaMa 13B). But doing so breaks compatibility for other LLaMa based models.

Relevant changes here.
https://github.com/nsarrazin/serge/blob/a837ea48e017289a21a9574b0fe862f541874a14/api/Dockerfile.api#L18-L20

We could make this more generic, but maybe it needs to be handled in llama.cpp instead ? Not sure yet.

[nsarrazin]

Looks like I spoke too soon, still need this fix for loading the 13B model.