Quarter turns and mirrors could avoid re-encoding entirely

A 90° rotation or a flip does not change a single pixel value, it only changes where they sit. We still send the whole image through renderToCanvas and re-encode it, so a JPEG loses quality every time someone straightens a photo, and a large one takes the full encode cost for a change a tag could express.

Asked for in nextcloud/viewer#2804, pointing at Memories which already does it.

The idea: when the only thing the state carries is some multiple of a quarter turn and/or a mirror, and the source is a JPEG, rewrite the EXIF orientation tag on the original bytes and hand those back untouched.

• Recognise the case, no crop, no adjustment, no filter, no annotation, and rotation a multiple of 90
• Map our rotation and flip state onto the eight EXIF orientation values
• Return the original bytes with the tag rewritten
• Fall back to the normal path for everything else, including PNG and WebP where there is no tag to lean on

Depends on #18, it needs the same reader and writer. Worth doing after it rather than inventing a second half-parser.

👾 Found while going through the viewer's editor issues with Claude Code.

Start at the existing renderToCanvas path and read the shared reader and writer from #18. The work is done when eligible JPEG quarter-turn and mirror states return the original bytes with the correct EXIF orientation, while other edits and PNG or WebP continue through the normal path.