Support Metal inference

[jonfairbanks]

The latest version of llama.cpp has huge improvements on Mac hardware.

Could this be implemented here as well?

ggerganov/llama.cpp#1642

[louisgv]

Yup, it's based on top of ggml and although the rust implementation is a slight deviation fro llama.cpp, it's likely portable. We can use Rust macros to deliver an aarch64 dedicated GPU inferencing as well

[LLukas22]

Since rustformers/llm@47a41c9 Metal support is in the main branch for llama based architectures, it can be enabled by compiling with the metal --features flag and by setting use-gpu to true in the model-config.