Re-evaluate Explicit SIMD Aggregations

[tustvold]

Is your feature request related to a problem or challenge? Please describe what you are trying to do.

We currently only have explicit SIMD kernels for min, max and sum. This is a non-trivial additional complexity, and relies on features not present in stable Rust.

Describe the solution you'd like

Historically LLVM has struggled with auto-vectorisation of horizontal operations like aggregates, and this has been why we've kept these kernels, however, it is possible this has since improved. We should assess whether it is possible to eliminate the need for explicit SIMD kernels, and potentially remove them

Describe alternatives you've considered

Additional context

#4560 contains some experiments for improving the auto-vectorisation of aggregates

[jhorstmann]

I started working on this and #5031 and think I have found a way to generalize the approach from #4560 and get auto-vectorization also for MIN/MAX. Still need to clean up the code further and run more benchmarks.

[jhorstmann]

@tustvold I must have missed or forgotten that the compare kernels and arithmetic div/mod also no longer use explicit simd. Just to confirm, all the methods in ArrowNumericType behind the simd feature would no longer be needed once aggregation stops using them?

[tustvold]

Correct, it would definitely simplify things to no longer have nightly only functionality

[tustvold]

label_issue.py automatically added labels {'arrow'} from #5100