ggml : increase GGML_MAX_DIMS

[ggerganov]

The 3D convolution operator requires 5D tensors:

https://pytorch.org/docs/stable/generated/torch.nn.Conv3d.html

If we cannot figure out a workaround (e.g. using a combination of ggml_view + ggml_reshape + ggml_conv), we might have to increase GGML_MAX_DIMS. Not sure how difficult this change would be. At the moment, it is not a priority (currently needed by the sd.cpp project leejet/stable-diffusion.cpp#491 (comment)), but it's something we should keep in mind.

[JohannesGaessler]

If we do this it would maybe make sense to add something like ggml_backend_op_max_supported_dim to prevent the misuse of high-dimensional tensors without having to touch every single op in a single PR.

[JohannesGaessler]

Although on second thought you could also just bake this logic directly into ggml_backend_op_supports_op.