Use KernelAbstractions backend

[charleskawczynski]

This is a big step towards closing #32.

What else is needed:

• Remove CUDA/Adapt extension/dependency
• Add other backend tests?
• Improve the performance
• Fix Add CUDA tests #34

Right now, at least on the CPU, the KA kernels are more than 5x slower. cc @avik-pal @vchuravy

[avik-pal]

Thanks for looking into this so quickly. I generally tend not to use KA for CPU (in favor of just the @simd loop), but I think it uses threading for CPU, which might not be ideal for smaller broadcasting problems.

[charleskawczynski]

Thanks for looking into this so quickly. I generally tend not to use KA for CPU (in favor of just the @simd loop), but I think it uses threading for CPU, which might not be ideal for smaller broadcasting problems.

Yeah, I thought I'd give it a try, it's kind of a nice solution, to be honest. The only reason CUDA is still needed in the extension is for Adapt. I think that this branch should mostly support other backends, but if Adapt is needed for other backends, then I think you'd need to add it. I don't have experience adapting AMD gpus, but I imagine the pattern is very similar.

Feel free to try things on this branch. I probably won't merge this in until we add tests (to close #34), and perhaps fix the cpu perf, or just revert things to how we previously had it.