Extending scatter! to work with CUDA sparse arrays

[alonsoC1s]

Aims to fix #647 by extending 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. In my testing the existing CUDA kernels work perfectly fine with CuSparseArrayCSC.

The proposed implementation, perhaps inelegantly, just expands the types in the signature with Union{...}. I am open to discussing more beautiful ways of implementing this. Ideally, AbstractCuSparseArray would be a subtype of AnyCuArray.

PR Checklist

• Tests are added
• Documentation, if applicable

[alonsoC1s]

The integration test with Lux fail because always_inliner! is not defined, and what look like Enzyme internal errors. Not sure if this is unrelated

[mcabbott]

Seems fine. Is it possible to add a test on CI somehow, perhaps in https://github.com/FluxML/NNlib.jl/blob/master/test/ext_cuda/scatter.jl ?