feat: cuda implementation for ggml_conv_transpose_1d

[balisujohn]

adds a cuda implementation for ggml_conv_transpose_1d

may extend the CPU version of the op to allow padding (currently, padding must be zero).

Currently a draft, posting now to prevent duplication of effort, I will remove draft status when ready for review.

[balisujohn]

Alright, this is ready for review. The key questions/concerns I have are as follows:

I added some tests for ggml_conv_transpose_1d in test-conv-transpose-1d.cpp some of these tests mirror the conv1 1d transpose tests in test-conv-transpose.c I have verified that the tests in test-conv-transpose-1d.cpp pass in both cuda and CPU mode, but I'm not sure if what's currently in the PR will run the tests in CUDA mode in the CI/CD, I had to remove the manual #define GGML_USE_CUBLASI had added in order for the CI/CD build to work. Curious how best to combine the new tests I added with what's already there, and how to make sure the tests I added are actually running with CUDA in the CI/CD process.

There is an extremely bizarre problem I'm encountering when trying to use this op with tortoise, the result is all zeros, but if this line is changed from

    dst[global_index] = accumulator;

to

    dst[global_index] += accumulator;

, the result is no longer all zeros.
I'm going to try to isolate the error into a self-contained example in a ggml fork so other people can take a look at it.

Some more background, the last test in test-conv-transpose-1d.cpp uses tensors that are the same shape and calls the op with the same params as the one with the zero behavior in tortoise, and seems to work correctly, so I don't think it's just because of the shape of the tensors and parameters to the op. It seems like src0 and src1 are pointing to zero valued vectors in this case, so even though the logic is correct the accumulator is being set to zero. And for some reason dst is pre-populated with non-zero values, so both += and no assignment to dst result in nonzero values being present in the result.
(The weird bug is fixed)

[slaren]

The github runners cannot run CUDA code. There is the ggml-ci can run CUDA tests, but it only monitors branches in this repository. A test should be added to test-backend-ops as well.

[forworldm]

int -> float

[balisujohn]

yeah that seems to fix the issue I was experiencing; it's bizarre that the tests still passed even in cuda mode with the types accidentally set to int Thanks so much!

[forworldm]

int -> float

[balisujohn]

I fixed the accumulator bug, added a test to test-backend-ops; ready for review again :^)

[ggerganov]

Also assert contiguous src0 and src1

[ggerganov]

I don't see update of ggml_backend_cuda_supports_op() to indicate that this operation is now supported. While at it, update the rest of the GPU backends to return false in their corresponding supports_op() functions so that the tests do not fail

Maybe add a few more small tests to tests-backend-ops just in case