It's useful to set a slice size != border-width, as backgrounds are
clipped to border-width too.
As slices can be half-transparent and overlap the background,
this would not fill the border box properly if we only use a single
property for specifying the width.
Also, this brings us even closer to CSS3.
NULL is a valid value for border-image, so if it's NULL when unpacking,
don't try to access the struct fields, but just init the GValues for the
unpacked parameters.
The new code is smaller, less crashy and correct(er), but arguably more
complex. I'd have liked to make it simpler, but this border image
algorithm is complex...
It's not useful to cache these surfaces here, as the GtkBorderImage will
be always generated on the fly, being a shorthand property.
This also allows to get rid of the intermediate image surfaces for
rendering the slices; we now use cairo_surface_create_for_rectangle()
to proxy the slices from the source surface to the rendered area, which
should also yield better performance.
