feat: add dizzy dithering (DitherMode 9)

[g4bri3lDev]

Adds dizzy dithering — error diffusion that visits pixels in pseudo-random order instead of a raster scan, spreading each pixel's error only to neighbours not yet quantized. No scan direction means no directional artifacts.

Algorithm by Liam Appelbe. Implemented from the written description — the article has no code listing, so this is an independent implementation, not a port.

opendisplay_structs.h:187 states DitherMode ids are owned by this repo, so no protocol change is needed; 9 was the next free value.

Evaluation — read this before merging

The article demonstrates the technique on continuous-tone greyscale. This library targets 2–7 ink measured palettes where hue error dominates, so the result does not transfer cleanly. Mean OKLab ΔE on 4×4-block-averaged output, 8 fixture images:

mode	mean ΔE
burkes	0.1520
dizzy	0.1528
floyd_steinberg	0.1535

Overall it is a wash — the spread is within noise. Split by palette it is not:

• spectra6 (the measured 6-ink palette this library actually targets): dizzy wins or ties on 7/8 images, up to ~17% better (unicorn.png 0.1321 vs 0.1596).
• mono (2-ink B/W): dizzy loses on 5/8 and is the worst of the three, up to ~14% worse (frankfurt_nacht.png 0.2762 vs 0.2425).

Metric sanity check: DitherMode::None scores 0.1843, ~21% worse than every diffusion mode — so the metric is measuring what we think it is.

Cost: 1.09× Burkes at 100×100, 1.25× at 400×300. The gap widens with size, consistent with random access over a multi-hundred-KB f64 buffer being cache-hostile. This matters because Home Assistant renders on this path.

Verdict: an opt-in mode for multi-ink palettes, not a new default. Burkes stays the default.

Correctness

The load-bearing property is that the traversal visits every pixel exactly once. If it didn't, pixels would be silently skipped or quantized twice and the output would still look entirely plausible — no image-quality test would catch it.

Multiplication by an odd number is invertible mod 2^k and XOR is self-inverse, so five rounds compose to a bijection. Verified exhaustively for all n from 1 to 3000 plus real panel sizes (384,000 / 524,288 / 1024×576): zero failures.

Also pinned: error-conservation over all 255 non-empty neighbour subsets, determinism, canonical-ink pinning, and a frozen 16-value vector proving Rust, Python and TypeScript agree byte for byte. 16 new regression references (8 images × 2 tags) bring the suite to 40; zero existing references changed.

Notes

• serpentine is ignored — there is no scan direction to reverse. Documented alongside NONE and ORDERED.
• Canonical-pinned pixels are marked processed before the walk, so no neighbour wastes error on a pixel that will ignore it. This deliberately differs from the raster path, where pinned pixels accumulate error that is then discarded.
• When every neighbour is already quantized the error is dropped. Inherent to the algorithm, and commented so nobody "fixes" it.
• Pre-existing bug found, not fixed here: benches/dithering.rs's bench_real_images unconditionally loads a missing katzi.png at group-construction time, so an unfiltered cargo bench panics. Confirmed present on main. Worth a separate PR.
• Downstream: odl-renderer and py-opendisplay must adopt id 9 before this is reachable from Home Assistant.