Stop relying on backend type to determine algorithm used

[christiangnrd]

This is non-breaking, since the default behaviour remains unchanged. However, use_KA = true with cpu arrays will not work with KA <0.10.

Also includes some fixups for things I missed in #39. I can split that off into a separate PR if more time is needed to review the main changes here.

[anicusan]

This is so, so useful, thank you @christiangnrd ! My only "issue" is aesthetic with the kwarg name - "use_KA" is not lowercase snake_case 😅. What do you think about something like "prefer_threads=true"? I think "prefer" makes more sense, as you still have an escape hatch if you wrap your array into a CLArray.

[christiangnrd]

I was hoping you would have a suggestion for a better name! I'll make the change at some point in the coming days.

[christiangnrd]

Not sure I understand why reductions fail with the KA pocl backend, but not the OpenCL pocl backend.

Maybe we merge but don't document/support prefer_threads until KA 0.10 is released and tests pass?

[christiangnrd]

@anicusan I've made the switch to prefer_threads, but I also removed it from the docstrings since reductions don't currently work with KA's 0.10 pocl backend. This makes this PR technically a non-functional change, but the repo is more ready for the eventual removal of GPU <: Backend in KA.jl.