Testing currently not possible with FreeImage, to be done when migrating
to OpenImageIO as part of: #3459

…e sizes
The layered array's row pitch is the maximum width over all source
images, but since the OpenImageIO migration (colmap#3459) each bitmap was
copied with a single contiguous memcpy, shearing every source image
narrower than the widest one. Restore the row-by-row strided copy,
mirroring the source depth map upload below.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The layered texture reads each layer with a row stride of max_width,
while the bitmap rows are packed with the image's own width. Since
commit  8e014c5  (colmap#3459) the bitmaps were copied with a single
contiguous memcpy, displacing row r of every image narrower than
max_width by r * (max_width - width) pixels. Example for a 2x2 image
in a 3-wide layer:
  bitmap [a b; c d] -> contiguous copy [a b c; d 0 0] (wrong)
                       row-wise copy   [a b 0; c d 0] (correct)
Workspaces where all source images share the same dimensions are
unaffected (the two copies coincide there), which is why the standard
single-camera case never showed the problem. Workspaces with mixed
image sizes get silently corrupted photometric costs for every source
image narrower than the widest one.
Restore the row-by-row copy used up to commit  7d2058b  (the parent of
 8e014c5 ), mirroring the source depth map upload below.