feat: adds ggml_pad_ext to allow prefix padding

[balisujohn]

modifies ggml_pad so you can specify beginning and end padding for all 4 dimensions of a tensor. Previously, padding could only be added at the end of the first 2 dimensions. The padding is still constant padding with a value of zero, as before.

Adds new test file test-pad.cpp

[balisujohn]

I forgot that ggml_pad actually supports 4 dimensions as-is, I will add the support for pre-fix padding in all 4 dimensions, so marking as a draft for now.

[balisujohn]

Alright so it supports up to 4d prefix and postfix padding with the CPU version. for cuda it supports up to 3d prefix and postfix padding (the cuda pad kernel was previously limited to 3d tensors and I didn't remove this limitation). Two of the tests in test-pad.cpp use 4d tensors so they would need to be commented out to run the tests in cuda mode, though I verified all the other tests work with the cuda kernel.

This modification leaves the arguments of ggml_pad as-is, and adds ggml_pad_ext which requires the desired prefix and postfix length to be specified for all 4 dimensions.

Ready for review :^)

[balisujohn]

(I noticed I need to change some variable names in test-pad.cpp and there are similar minor fixes needed in the pad_reflect_1d PR)

[balisujohn]

this and pad_reflect_1d are ready for review again :^)