Add merge and merge_n algorithms

[pepijnve]

Which issue does this PR close?

• Closes Provide algorithm that allows zipping arrays whose values are not prealigned #8752.

Rationale for this change

The algorithms suggested in this PR originate from the case logic in DataFusion (see datafusion#18152 and datafusion#18444). I think it might be useful to move them to arrow-rs instead of being tucked away in a corner of the DataFusion codebase.

What changes are included in this PR?

Adds a two-way and n-way merge algorithm that's halfway between zip and interleave. In contrast to zip the truthy and falsy arrays do not need to be prealigned. In contrast to interleave the relative order of elements in each input array is retained in the final result.

Are these changes tested?

I've already added two minimal unit tests, more should probably be added.

Are there any user-facing changes?

No breaking API changes

[pepijnve]

The optimisation work that was done in #8653 would make sense here as well. That has not been done yet.

[alamb]

Thanks @pepijnve -- what do you think about also adding benchmarks to this kernel (so that future optimizations work better)

[pepijnve]

what do you think about also adding benchmarks to this kernel

Good idea. I’m happy to continue working on this one. I created the PR already to get the ball rolling and solicit input from other devs.

[pepijnve]

The optimisation work that was done in #8653 would make sense here as well. That has not been done yet.

While looking into this I realised that merge on scalars is effectively identical to zip so I resolved this by delegating to zip in case of scalar input

[pepijnve]

what do you think about also adding benchmarks to this kernel

@alamb I duplicated the microbenchmark for zip as a quick fix. Is it worth trying to actually share the sets of input data and masks? If so, where should I move that code?