sync: update ggml

[stduhpf]

Mostly fixes and optimisations, especially for Vulkan backends.
(Fixes #439 for example)

[stduhpf]

Oh, I think this breaks Q2_K and Q3_K quants on Vulkan. At least on my GPU.
Reverting ggerganov/ggml@941c450 fixes it for me.

[daniandtheweb]

I've tried applying these changes and syncing ggml to the latest commit and vulkan image generation seems to be broken. I think it may be worth taking a look at it, some of the latest ggml changes supposedly greatly improve stable-diffusion.cpp's performance.

[daniandtheweb]

The issue that this sync has is that the main program can't autodetect the weight type anymore.
When manually specifying the type the program runs as expected (on my card the speed for 512x512 sd 1.5 goes from 1.37 it/s to 1.44 it/s with the latest changes on upstream ggml).

[stduhpf]

Yeah this PR isn't up to date with the lastest GGML changes. ggml types have been changed again ,so enum sd_type_t in stable-diffusion.h needs to be updated for it to work with lastest GGML.