perf: AVX2 gather for PixmapRef::gather

[bhark]

This PR is part of a series linked to #174.

What

Replaces 8 scalar pixel loads in PixmapRef::gather with _mm256_i32gather_epi32 (on AVX2). Again gated on feature = "simd", target_feature = "avx2", meaning SSE2/SSE4.1/AVX, Neon etc. are unchanged.

Why

Image sampling takes a hard hit from gather, because bicubic patterns make 16 sample calls per pix batch, each loading 8 pixels. One AVX2 gather replaces 8 dependent scalar loads.

Results

	speedup
Geomean (56 benches)	1.04x
patterns::hq (bicubic)	1.44x
patterns::lq (bilinear)	1.33x
patterns::plain (nearest)	1.18x
fill::rect, hairline, blend::destination_over (don't hit gather)	1.00x

No regressions. A bunch of unrelated benches landed at ~1.1x, probably just noise/variance.

Notes

• _mm256_i32gather_epi32 does no bounds check; callers camp indices to [0, w*h] so the invariant holds
• Style-wise inspired by existing src/wide/i32x8_t.rs / u32x8_t.rs
• Only seen when built with -Ctarget-cpu=haswell, -Ctarget-feature=+avx2 or target-cpu=native
• Benched on i5-13400F (raptor lake). Since gather is microcoded, the win on Zen 1/2 may be smaller or nothing at all. Still shouldn't regress.

[RazrFalcon]

Looks good, but again, we should wrap it nicely in wide.