Update to the latest llama.cpp ggml

[philpax]

Another day, another suite of changes. Since our last update, M1 and AVX inference has improved, as well as a bunch of other improvements: https://github.com/ggerganov/llama.cpp/compare/437e77855a54e69c86fe03bc501f63d9a3fddb0e..HEAD

If you're interested in tackling this, check out our contributing document: https://github.com/rustformers/llama-rs/blob/main/CONTRIBUTING.md

[philpax]

Note that this changes the implementation of the LLaMA model - it is not a drop-in upgrade!

[KerfuffleV2]

I'm working on this one. Just updating the bindings seemed pretty uneventful: ~~~some types changed from i64 to i32. After that fix it compiles and appears to function.~~~ edit: Killl meeee, I accidentally copied from the wrong llama.cpp directory and ended up reverted to an older version...

Still need to do the actual model change part, which hopefully will put #67 to bed in a permanent way.