perf(kpm): SIMD path for Pyramid downsample (AVX2/SSE4.1/wasm32)

[kalwalt]

Summary

Add SIMD implementations of kpm::freak::pyramid::downsample gated behind the project's existing feature flags, with criterion benchmarks proving the speedup over the scalar baseline.

Context

Follow-up to #130 (M8 step 1). The PR landed a clean scalar baseline. The downsample inner loop is a textbook SIMD target — for each output pixel it reads 4 source pixels and computes (sum + 2) >> 2.

The project already has feature flags ready for this:

• simd-x86-avx2
• simd-x86-sse41
• simd-wasm32

Canonical pattern from CLAUDE.md §3:

#[cfg(all(target_arch = "x86_64", feature = "simd-x86-avx2"))]
if is_x86_feature_detected!("avx2") {
    return unsafe { avx2_impl(input) };
}
// scalar fallback
scalar_impl(input)

Depends on

• perf(kpm): add criterion benchmark for Pyramid downsample #131 (benchmark must land first to establish the baseline this PR will beat)

Scope

• Implement downsample_avx2 (process 32 output bytes per iteration via 256-bit lanes).
• Implement downsample_sse41 (16 output bytes per iteration via 128-bit lanes).
• Implement downsample_wasm32 using core::arch::wasm32 SIMD intrinsics.
• Each path behind unsafe + cfg(target_arch) + cfg(feature = ...) + runtime detection (is_x86_feature_detected!).
• Each unsafe block prefixed with a // SAFETY: comment explaining the invariant.
• Output must be byte-identical to scalar — add a property test that compares scalar vs SIMD on random inputs.

Acceptance criteria

• Output of each SIMD path is byte-identical to the scalar baseline for all tested input sizes (including odd dimensions and small images).
• Each SIMD path shows a measurable speedup over scalar on its target architecture, reported in the PR description with concrete ns/iter numbers from perf(kpm): add criterion benchmark for Pyramid downsample #131's benchmark.
• All existing kpm::freak::pyramid tests still pass.
• cargo clippy --all-targets --all-features clean.

Out of scope

• rayon parallelization (could be a separate issue or bundled here — open question).
• Chunked memory layout (perf(kpm): SIMD path for Pyramid downsample (AVX2/SSE4.1/wasm32) #132).

References

• PR feat(kpm): port BoxFilterPyramid8u to Rust (M8 step 1) #130 — original port
• perf(kpm): add criterion benchmark for Pyramid downsample #131 — benchmark baseline
• refs feat(kpm): M8 — FREAK descriptor & DoG detector in pure Rust #125 feat(kpm): M8-1 - port pyramid.h + pyramid.cpp to freak/pyramid.rs #126 — M8 parent issues
• CLAUDE.md §3 — SIMD conventions