ggml_compute_forward_concat() for arbitrary tensor type

[vmobilis]

It is needed for embeddings in stable-diffusion.cpp when weight type is not f32.

And thanks for the program! )

[vmobilis]

@ggerganov, it seems that my patch is wrong, the correct one is leejet/stable-diffusion.cpp#601.

[vmobilis]

On second thought, @ggerganov, maybe support of different types is still usable?

[vmobilis]

On 3rd thought, it doesn't work after all... 🙃
At least for someone: leejet/stable-diffusion.cpp#601 (comment)

[vmobilis]

@ggerganov, is anything wrong with commit now?

[ggerganov]

We can merge - I thought that something was wrong based on your last comment so didn't review.

[vmobilis]

@ggerganov, thank you, black images were generated even without updated concat(), and

ggml/src/ggml-cpu/ggml-cpu.c

Line 6708 in c6a6a64

GGML_ABORT("fatal error");

was not called when generating (and patch avoids exactly that break), therefore the error is probably somewhere else and is unrelated to commit.