Extending scatter! to work with CUDA sparse arrays (retry of #648)

[alonsoC1s]

This is a new attempt at PR #648

It extends the signature of scatter! to work with AbstractCuSparseArray, a CUDA array type notably excluded by the original method. With the proposed patch, calling scatter! with sparse arrays from CUDA.CUSPARSE will correctly call the CUDA-specialized method instead of calling the generic CPU method, which triggered a scalar indexing error.

[alonsoC1s]

Quick update. The test failures were significant, and I underestimated what needed to be done. I'll mark the PR as draft while I work on JuliaGPU/GPUArrays.jl#694, which will hopefully make finishing this PR for all of the different sparse formats more feasible

[alonsoC1s]

The implementation of scatter! and friends should now work with GPU sparse device arrays as a source, though not as a destination, since that would potentially require allocating in device code. I recreated the standard testing suite and specialized it so that it makes sense in the sparse case

This PR is heavily dependant on a related PR to GPUArrays.jl being merged, and shouldn't make its way into main until the other one is succesfully merged and a new release is made

[alonsoC1s]

Looking at the CI failures, I can't tell if they are related. They seem to happen only when doing gradients with Enzyme, and some don't even fail for scatter!, but other functions. While testing locally, it all worked as expected