nanosvg walks every scanline of the target for each visible shape, so
rasterizing the whole SVG and cropping the images out of it cost
shapes * canvas area. Each image is now rasterized into a buffer the size
of its bounding box, from a list of copies of its own shapes, so the
images can be rasterized on the thread pool.
Atlas generation on a Redmi 9A (32-bit) goes from about 165 ms to about
95 ms, and on a desktop from about 10 ms to 6 ms. Parsing the SVG is now
the largest part. The atlas layout is unchanged; a few edge pixels differ
by at most 2/255 from the different subpixel offsets.
Also copy whole rows in Image::copyfrom, and fix a leak of the SVG file
data when parsing fails.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
