Use KernelAbstractions.jl for gather/scatter kernels

[pxl-th]

Depends on #486.
Successfully tested on CPU, CUDABackend, ROCBackend.

Since LLVM does not support certain atomic operations (fmin, fmax, fmul, fdiv) tests for them are disabled when using such backends, which is currently only ROCBackend:
https://llvm.org/docs/LangRef.html#atomicrmw-instruction

fmin/fmax requires LLVM 15+: https://reviews.llvm.org/D127041

TODO

• Add @inbounds.

PR Checklist

• Tests are added
• Documentation, if applicable

[pxl-th]

It may be that LLVM on Julia 1.6 is too old and we can't support all of the ops that we currently support for CUDA.
On 1.9-rc2 it fails for - op, but all other ops are working...

As an alternative, we can retain (for now) NNlibCUDA scatter.

[pxl-th]

I've retained scatter kernels for CUDA in NNlibCUDA so that this PR is not blocked.
But what's important is that other backends that rely on KernelAbstractions.jl will be able to use embedding layers (fwd + bwd pass), which is probably the most common use-case for it.

[ToucheSir]

FWIW the copy of NNlibCUDA in this repo doesn't do anything at present, so we need to make sure the overloads in the standalone NNlibCUDA repo still work.

[pxl-th]

FWIW the copy of NNlibCUDA in this repo doesn't do anything at present, so we need to make sure the overloads in the standalone NNlibCUDA repo still work.

IIC, if NNlibCUDA in this repo is not registered, then tests at using NNlibCUDA line should import the standalone repo.
And I've disabled Gather and Scatter tests from test suite for CUDA, because we call them from NNlibCUDA tests.

[pxl-th]

Not sure what's with CUDA on 1.8...
Will investigate in a bit

[pxl-th]

Probably was some kind of hiccup, now the CI passes.

[ToucheSir]

Anything left before merging?

[pxl-th]

Anything left before merging?

I think no, should be good to go