perf(compute): vectorize mixed int32 filters with AVX2

[fallintoplace]

Summary

• Add an AVX2 path for filtering 32-bit fixed-width values with mixed selection masks.
• Compress two groups of four values with vpshufb and use masked stores for the compact output.
• Keep the current path for nulls, short or unaligned inputs, dense masks, unsupported CPUs, and noasm builds.
• Add tests and a benchmark for mixed masks, offsets, and 1K, 64K, and 1M rows.

Benchmark

The benchmark is in arrow/compute/vector_selection_bench_test.go.

It covers alternating, random 25/50/75%, clustered 50%, all-selected, and none-selected masks at 1K, 64K, and 1M rows.

A native AVX2 run needs an x86_64 machine with AVX2. The benchmark also runs on the local Apple M1, but that only exercises the existing arm64 fallback, so no local AVX2 timing is included.

Tests

• go test ./arrow/compute/internal/kernels ./arrow/compute
• go test -tags noasm ./arrow/compute/internal/kernels ./arrow/compute
• go test -race ./arrow/compute/internal/kernels ./arrow/compute
• amd64 cross-build for both packages

[zeroshade]

Dispatch guards are sound (AVX2 probe, length%8, byte-aligned filter offset, bounds check, null-free inputs only, mixed-mask heuristic) and GetSpanValues applies span offsets, so the sliced-input case is safe. The c2goasm source is included per repo convention. The AVX2 path itself will be exercised by the AMD64 CI jobs I've just approved.

---

This review was drafted by an AI-assisted tool and
confirmed by an Apache Arrow Go maintainer.