Backport the performance improvement from llama.cpp

[kovaacs]

It would be very cool if the performance improvements from ggml-org/llama.cpp#613 could be backported to this repo.

I couldn't find an issue for this, if there is one, I'm happy to close this.

[edwios]

Related. #702

Since whisper.cpp shares the same llama.* and ggml.* files with llama.cpp, I think it is just one step away - convert the whisper model from ggml into the new ggjt format.

[ggerganov]

Just updated ggml and the performance should be much better.
I observe > x2 overall improvement on Apple Silicon

No need to generate new models - everything should just work