Dependency discussion: portable SIMD crate (wide/pulp) vs the hand-rolled intrinsics in simd_scan.rs

Dependency discussion per CONTRIBUTING.md. Gated on the benchmark-harness issue — this should
not turn into a PR before there is a repeatable measurement.

What exists today

Hand-rolled, and deliberately minimal:

• src/align/simd_scan.rs — find_stop, a 16-byte chunked "does this chunk contain a stop"
  reduction in SSE2 / NEON intrinsics, with a scalar re-scan to locate the exact position. The
  module comment states the design rationale: keep the arch-specific surface to a trivial boolean
  reduction, do position extraction in portable scalar code.
• src/cpu.rs — runtime feature detection, the build-flavour guard, and prefetch_read
  (_mm_prefetch / prfm).

That is a good design and the bar for replacing it is correspondingly high.

Candidates

The actual questions

1. Does a portable crate buy wider vectors? Today's scan is fixed at 16 bytes, i.e. SSE2/NEON
   width. AVX2 (32 B) or AVX-512 would halve or quarter the chunk count, but only with runtime
   dispatch — which is exactly what wide does not do and pulp does.
2. Five platforms, including Windows. CONTRIBUTING.md calls this out. Any candidate has to
   build cleanly everywhere, including the aarch64 macOS path.
3. Results must be bit-identical across dispatch paths. A scalar fallback, an SSE2 path and an
   AVX2 path must return the same find_stop index. That needs a test exercising all compiled
   paths, not just the native one.
4. Is find_stop even hot enough to matter? Answer with a profile before touching it.

Checklist

• Profile find_stop's share of align wall time (blocked on the benchmark harness)
• Try widening the existing hand-rolled scan to AVX2 with is_x86_feature_detected! first,
  since that costs zero dependencies and answers question 1 directly
• Only if that is unsatisfying: prototype with pulp (multiversioning is the differentiator)
• Cross-path identity test: scalar vs each vector path
• Verify Windows + macOS aarch64 builds

Read CONTRIBUTING.md and the design comments in src/align/simd_scan.rs and src/cpu.rs. Wait for the benchmark harness, then profile find_stop, try the AVX2 widening experiment, and only prototype pulp if needed. Done means measured results, cross-path index identity tests, and verified Windows and aarch64 macOS builds.