Make quantization independent of the architecture (fixed histogram order, consistent NEON distance)

[dustinkirkland]

Fixes #130.

The same image quantizes to a different palette on x86_64 and aarch64 (details and measurements in #130). Two causes, three small changes:

1. Histogram built in a fixed order. Histogram::finalize_builder drained the colour HashMap in iteration order, and everything after that is order-sensitive (cluster bucketing by position, the f64 weight sum, median cut's tie-breaking, k-means' 256-item chunks). The hasher is deterministic, but hashbrown's probe-group width is 16 bytes with SSE2 on x86_64 and 8 bytes with NEON on aarch64, so colliding keys land in different buckets and the map iterates in a different order per architecture. Sort temp by colour before the cluster pass. One sort of at most max_histogram_entries items.
2. Posterize rehash in key order. init_posterize_bits re-inserts the old map with coarser keys; when several keys collapse, the last one inserted wins, which again depended on iteration order. Collect, sort by key, then extend. Last-wins semantics are kept (summing the boosts of collapsed entries would arguably be better, but that is a behaviour change and left for you to decide).
3. NEON f_pixel::diff adds in the same order as the other paths. The x86_64, scalar and WASM paths compute (r + g) + b; the NEON path used vpaddq_f32 and computed r + (g + b). Float addition is not associative, and the refinement loop amplifies the ulp differences in the metric into a different palette for some images (visible with --quality 0-100). Use the same association; it is also one instruction fewer.

Verification (pngquant 3.0.3 rebuilt with these changes, x86_64 native and aarch64 under qemu-user, comparing decoded pixels and palette sizes): before, 4 of 5 noto-emoji sample bitmaps differed between the architectures at --quality 85-95 and one more at --quality 0-100; after, all 11 cases (5 images × 2 quality settings, plus a 262,144-colour image at --speed 10 that goes through the posterize rehash) are identical on both. Each change was isolated: sorting the histogram alone fixed the 85-95 cases (and made the x86_64 output of the UN flag byte-for-pixel the previous aarch64 output); the NEON association alone explained the remaining 0-100 case (an x86_64 build with NEON's order reproduced the aarch64 output exactly). cargo test passes on both.

Output on a given architecture changes for images where either effect mattered: same quality target and algorithm, a fixed tie order and a consistent metric.