ollama: unvendor llama.cpp

[SomeoneSerge]

Issue description

Ollama vendors a copy of llama.cpp as a git submodule: https://github.com/ollama/ollama/tree/main/llm. Llama-cpp is built as part of the buildGoModule, apparently during the go generate. This significantly increases build times when iterating on the ollama derivation (because you have to rebuild the submodule each time you change the ambient ollama package). This also makes it much harder to configure (patch, tweak) llama.cpp used by ollama. Overall, I think we want less vendoring in Nixpkgs, we want fewer "build inside a build" situations.

It is desirable to refactor the ollama expression so as to reuse the prebuilt llama-cpp from Nixpkgs

EDIT: A detailed overview of the current situation is provided by @abysssol at #323249 (comment)

CC @abysssol @dit7ya @elohmeier @RoyDubnium

Historical context

This is a follow-up to #323249 and https://github.com/NixOS/nixpkgs/pull/323056/files#diff-99616cb74ed708e718f689a9fc8f62999fdff6164590233bbb1a81a0f8f8c08b, aiming to make Nixpkgs' ollama nicer and easier to maintain