Fix merge-sort when sub/work group size is small

[carbotaniuman]

On some implementations, the max subgroup size is 1, which is too small for the sorting algorithm to produce correct results. Make the value at least 4.

[dmitriy-sobolev]

@carbotaniuman, thanks for noting that issue. I would propose a slightly different approach correcting the chunk size, which appears to be more robust.

// The chunk size must not exceed two sorted sub-sequences to be merged,
// ensuring that at least one work-item processes them.
const _IndexT __chunk = std::min<_IndexT>(__is_cpu ? 32 : 4, __n_sorted * 2);

[carbotaniuman]

I tested the change on my local installation and it produces the correct results with the new change.

[dmitriy-sobolev]

LGTM, but I would ask @MikeDvorskiy to check the changes as well. I may be biased because I also contributed.

[MikeDvorskiy]

LGTM

[dmitriy-sobolev]

@carbotaniuman, thank you once again. If you want your contributions to be mentioned in CREDITS.txt, do not hesitate to create a PR with their brief descriptions targeting that file.