Use OKLab HyAB perceptual color difference

[mourner]

Closes #127. Replaces the perceptual color difference metric at the heart of pixelmatch — YIQ Euclidean distance (Kotsarenko & Ramos 2010) — with the vastly better HyAB distance in the OKLab color space (with toe-corrected lightness) (Ottosson 2020; Abasi et al. 2019).

The threshold semantics are preserved: distances are scaled so that 1.0 is the black↔white difference, and the default threshold: 0.1 is unchanged.

Why

• OKLab is a modern perceptual space fit to color-appearance data, with far better hue uniformity than YIQ (a 1960s NTSC transmission space never designed for measuring color difference).
• HyAB (|ΔL| + √(Δa² + Δb²)) is a city-block/Euclidean hybrid shown to track human judgment of large color differences better than pure Euclidean metrics — and large differences are exactly what a diff threshold operates on.
• The Lr toe fixes OKLab's one weakness for image diffing: its cube-root lightness over-expands near-black. The fix reduces sensitivity near black and increases sensitivity in midtones and highlights, which matches how humans perceive color difference better.

Behavioral changes

• Diffs are re-ranked, which is the point: pairs the old metric conflated are now ordered by perceptual difference. Some previously missed differences (notably hue shifts at similar luma) are now caught; some previously flagged imperceptible ones no longer are.
• Sampling color pairs that sat exactly on the old 0.1 boundary, their median distance under the new metric is ~0.12 — i.e. the same threshold value is on average ~20% more sensitive than before. Sensitivity is now nearly uniform across the lightness range (band medians 0.118–0.124), which the old metric never achieved.
• Test fixtures are regenerated accordingly. The AA detector is untouched and intentionally stays on gamma-space Rec.601 luma.

Performance

The new metric is more math per pixel (sRGB linearization, LMS transform, cube roots), offset by various performance optimizations:

• an sRGB→linear LUT and premultiplied LMS contribution tables (per-byte);
• an interpolated cube-root LUT (Math.cbrt is a slow library call);
• a specialized opaque-pixel path with the rare transparent path split out-of-line, keeping the dispatch inlineable;
• a 4096-entry direct-mapped cache of cube-rooted LMS + Lr values keyed by packed RGB (screenshots repeat colors heavily);
• the threshold test folded into the HyAB comparison so the sqrt is skipped entirely;

Net result on the benchmark: comparable to the old YIQ implementation despite the heavier color model.