AMD ROCm support with HIPBLAS

[xangelix]

ggerganov/llama.cpp#1087 was merged recently and should make its way into GGML shortly. I think (hopefully) it should require only minimal changes on this side.

[xangelix]

related to ggerganov/ggml#472

[LLukas22]

rustformers uses llama.cpp as it's ggml source, feel free to create an PR including this change, seams like you only need to adjust the build.rs of the ggml-sys create. I wont be able to test this as i don't own a amd gpu.