GH-50542: [C++] Fix ARROW_SIMD_LEVEL=NONE build, compile SSE4.2 kernels

[domibel]

Rationale for this change

Building with ARROW_SIMD_LEVEL=NONE and ARROW_RUNTIME_SIMD_LEVEL=MAX failed with:

bpacking.cc:39:49: error: 'unpack_sse4_2' is not a member of
'arrow::internal::bpacking'

The SSE4.2 dispatch table entries are emitted when either the compile-time or the runtime macro is set, but the kernels themselves were only compiled when the compile-time macro was set.

What changes are included in this PR?

I moved the byte_stream_split SSE4.2 instantiations into a dedicated translation unit, keeping the dispatcher's own unit at the baseline ISA.

Are these changes tested?

Yes

Are there any user-facing changes?

No

• GitHub Issue: [C++] Version 25.0.0 does not build with ARROW_SIMD_LEVEL=NONE #50542

[github-actions]

⚠️ GitHub issue #50542 has been automatically assigned in GitHub to PR creator.

[AntoinePrv]

The PR looks sensible but this may not be the only limitation. From what I understood when last working on this part of the code, SSE4.2 is mandatory when building Arrow C++.
I believe the reason at the time was the unconditional usage of popcnt, but perhaps now that we moved to std::popcnt this is more portable.
There may be more places in the CMakeLists.txt where the default compiler options do pick a minimal SSE4.2 baseline so that would be a place to look at as well.

In principle I think treating all SIMD as optional makes the code more uniform, and is a free abstraction since the dispatch will change to static when possible.

@domibel I'm curious what is your use case? Do you only wish to disable SIMD, or do you need to target below SSE4.2?

[domibel]

@AntoinePrv Debian is using the x86-64-v1 baseline, so we build with ARROW_SIMD_LEVEL=NONE but keep ARROW_RUNTIME_SIMD_LEVEL=MAX for dispatching at runtime. That means we don't disable SIMD, we just keep it out of the baseline. The PR is fixing a regression introduced somewhere after the 23 release.

@kou Could ARROW_SIMD_LEVEL=NONE be added to the nightly test coverage?