Add support for missing quants in CPY (Metal & CUDA).

[gcp]

Fixes #10976.

[ggerganov]

Could you make a separate PR just for the Metal changes? Btw, I think the copy kernels could be implemented by reusing the dequantize_qX_X functions, likely with a single template + 4 instantiations. Would result in much smaller code change and allows to generalize in the future to other quantizations.

[gcp]

Btw, I think the copy kernels could be implemented by reusing the dequantize_qX_X functions, likely with a single template + 4 instantiations. Would result in much smaller code change and allows to generalize in the future to other quantizations.

Reusing the dequantize_qX_Y functions works, but doing it with templates is a bit tricky because dequantize_q8_0 swizzles its results differently than all the others (boo!). It would've been nice if this had just been a ggml_get_to_fp32_cuda call but that doesn't deal with the permutations the CPY code is expected to handle 😢