feat: adds ggml_pad_reflect_1d

[balisujohn]

This adds an op that mirrors the behavior or PyTorch's ReflectionPad1d operation.

Implementations for cuda and CPU are provided.

Tests still need to be added, so marking as a draft PR from now.

The cpu version of the op was derived with permission from @PABannier's implementation discussed here: #819

[PABannier]

It looks great, @balisujohn. Thanks for starting the implementation! I'm wondering whether we should implement a ggml_pad_reflect and pass in an is_2D argument that handles the 1D and 2D cases as it is done for ggml_im2col.

[balisujohn]

I added a test in test-backend-ops and also added test-pad-reflect-1d.cpp which I verified works with and without the #define GGML_USE_CUBLAS, though I'm not sure if the cuda versions of the ops are currently tested in the CI/CD or how to accomplish that.

I think this is ready for review.

[balisujohn]

A test I wasn't sure how to add was "check if it correctly fails when called with a pad length not shorter than ne0 of the input tensor."

[balisujohn]

Or and wrt the is_2D argument I support it but I'd ideally like to leave it to a future PR.

[ggerganov]

Some similar comments as in the ggml_conv_transpose_1d PR

Regarding the is_2D: currently the im2col is an exception - the only op that uses this pattern. Probably it is better to change im2col to not use it, since the alternative pattern with dimensions suffix (i.e. _1d, _2d, etc.) is well established

[balisujohn]

OK ready for review again : ^)

[PABannier]

Hello @slaren @ggerganov ! Do you plan to merge this operation? This would be super helpful for Encodec.cpp to support the Metal backend.

[ggerganov]

Hey @PABannier, this PR needs to be updated to the latest ggml changes before merging.

[PABannier]

Hello @balisujohn ! Do you plan to finish this PR soon? Otherwise, I'm happy to finish yours if you give me the permission to push on your fork.