buffer: reduce unnecessary painting to new cairo surfaces

Add buffer_adopt_cairo_surface(), which allows wrapping an existing
cairo image surface in a struct lab_data_buffer. This is useful when
loading PNGs since most will be loaded as ARGB32 already.

Fix a memory leak in the non-ARGB32 PNG case, where we do still need to
paint to a new image surface -- we were leaking the original surface.

Eliminate an unnecessary temporary image surface in SVG loading and just
render the SVG to the image surface held by the lab_data_buffer.

I also cleaned up and clarified the buffer API a bit:

- Add a pointer to the held cairo_surface_t (so we can still access it
  if there is no cairo_t).
- Remove the free_on_destroy bool (it was always true).
- Rename unscaled_width/height to logical_width/height and add an
  explanatory comment. It was unclear what "unscaled" meant.
- Rename buffer_create_wrap() to buffer_create_from_data().

This is laying groundwork for some more icon fixes I am working on
(making sure icons are loaded and rendered at the correct scale).

src/ssd/ssd-shadow.c

@@ -196,8 +196,8 @@ set_shadow_geometry(struct ssd *ssd)
 		 * is different).  The buffers are square so width == height.
 		 */
 		int corner_size = active
-			? theme->shadow_corner_top_active->unscaled_height
-			: theme->shadow_corner_top_inactive->unscaled_height;
+			? theme->shadow_corner_top_active->logical_height
+			: theme->shadow_corner_top_inactive->logical_height;
 
 		wl_list_for_each(part, &subtree->parts, link) {
 			set_shadow_part_geometry(part, width, height,