Non-mutually exclusive features

[pedro-devv]

At the moment GPU features are all mutually exclusive. Users of the library should be able to compile all backends they wish to, and then decide which to use at runtime. This would be more convenient for shipping programs using this crate.

One initial idea is to compile different Llama.cpp versions, one for each backend, as a dynamic library, however build artifacts aren't directly included in a project unless they are somehow included in a Rust source file, so if a project is bundled, the dynamic libraries wouldn't be included.