Vulkan: Fix Deepseek V2 inference by making ggml_vk_op_supports_incontiguous(GGML_OP_RMS_NORM) return true

[stduhpf]

Fixes #12956

It seems to work perfectly fine so far, but i don't really know what I'm doing, so maybe it's actually not supported and this is breaking something else.

(Also maybe the same should be done for GGML_OP_RMS_NORM_BACK?)

[jeffbolznv]

I don't think the shader actually supports noncontiguous tensors because it doesn't take into account the strides the way the cuda kernel does. If this is supported, we should also be able to return true from ggml_backend_vk_device_supports_op, but when I do that it fails the backend tests.

[stduhpf]

I see I was kinda expecting that kind of problem. Still, doing this workaround seems fixes deepseek v2 inference on my machine.

[0cc4m]

We would have to modify the shader to allow noncontiguous tensors, then.

[jeffbolznv]

@stduhpf do you want to try to make the shader changes yourself? If not, I can probably do it soon.

[stduhpf]

I don't think I have the skills to do it myself, I can give it a try, but I will probably not be able to figure it out.