[toddbeauchamp]

Taude:

Follow-up to #19 (stacked on it): large images are now shrunk off the UI thread, so the first frame that draws a big avatar no longer drops.

• Before: Window::paint_image box-filtered an oversized image inline, inside the atlas upload, the first time it was drawn at each downscale level. That was 11 ms for a 2048px avatar on an M5 Max, a dropped frame at 60 Hz, and slower Windows machines are worse.
• Now: images over 128x128 pixels are shrunk on the background executor, every frame of the image in one task. Until the copy lands, paint draws the nearest copy already in the atlas (so a resize keeps showing the old size), or nothing on first sight. The views that painted the image are notified when it lands, so cached views repaint too.
• Small images (128x128 and under, about 10 µs) are still shrunk during paint, so icons never pop in.
• Faster filter: fixed 2x2 blocks over chunked rows, and no per-channel division for fully opaque blocks. Output is bit-identical to Downscale oversized images before atlas upload #19's filter.
• Unchanged: the downscale_level cache key, drop_image freeing every level (it now also cancels a pending copy), and has_image_atlas_entry reporting what is in the atlas.

Timing

Filter time, release build, M5 Max, best of 30, shrinking to 32px:

Source	#19 (on the UI thread)	This PR (off the UI thread)
512px	0.78 ms	0.14 ms
1024px	2.72 ms	0.49 ms
2048px	10.98 ms	1.99 ms

Images with transparency run about 30% slower than opaque ones on the new path (2.66 ms at 2048px). Paint itself now only does an atlas lookup and a task spawn for these images.

Testing

• New elements::img tests: first paint draws nothing and the next frame after the task draws a 32px tile; a resize keeps the 256px copy until the 32px one is ready; drop_image cancels a pending copy; an animated image gets every frame's copy at once; a 128px image is still shrunk during paint; a cached child view repaints when its copy lands (fails if the notify is removed).
• Old and new filters compared on random opaque and translucent images, including odd and 1px sides, at levels 0 to 7: identical output.
• cargo test --workspace, cargo clippy --workspace --all-targets and cargo fmt --all --check pass on macOS.
• The Hub's headless visual capture draws 3 frames with run_until_parked between them, which covers decode plus shrink, so badges and avatars still appear in captures.