Performance: SIMD acceleration for hot paths (optional)

[enzoftware]

Description

Optimize performance-critical filter operations using SIMD intrinsics where beneficial. This is an optional enhancement for production readiness.

Context

Convolution-based filters (Gaussian, Sobel, Box blur) and pixel-level effects (brightness, contrast, grayscale) are prime candidates for SIMD optimization.

Tasks

• Profile and identify the top bottleneck operations using criterion benchmarks
• Evaluate using std::simd (nightly) vs packed_simd2 vs manual intrinsics
• Implement SIMD versions for the highest-impact operations
• Maintain scalar fallback for targets without SIMD support
• Benchmark before/after to quantify improvement

Acceptance Criteria

• At least 2x speedup on SIMD-optimized operations (measured via benchmarks)
• Scalar fallback works correctly on all targets
• No correctness regressions (same output as scalar version)
• SIMD code is behind feature flags if using nightly features