We don't need to read the embedded palette information with the rest of
the header, we can defer that to much later when the palette is actually
needed. This makes more sense conceptually, simplifies the code a bit,
and saves some bytes in the image state structure.
Since we're basically rewriting lbx_img_getpalette anyway, take this
opportunity to give it a better interface. It now returns the number
of embedded palette entries, allowing us to eliminate the palette size
feature in lbx_img_getinfo.