Different bits of code were using slightly different checks to test
whether a window was an attached dialog. Add a new
meta_window_is_attached_dialog(), and use that everywhere.

Also, freeze the is-attached status when the window is first shown,
rather than recomputing it each time the caller asks, since this could
cause problems if a window changes its type after it has already been
attached, etc. However, if an attached window's parent is destroyed,
or an attached window changes its transient-for, then fix things up by
destroying the old MetaWindow and creating a new one (causing
compositor unmap and map events to be fired off, allowing the display
of the window to be fixed up).

Remove some code in display.c that tried to fix existing windows if
the gconf setting changed, but which didn't actually do anything (at
least under gnome-shell). However, if 654643 was fixed then the new
behavior with this patch would be that changing the gconf setting
would affect new dialogs, but not existing ones.

https://bugzilla.gnome.org/show_bug.cgi?id=646761

An ARGB window with a frame is likely something like a transparent
terminal. It looks awful (and breaks transparency) to draw a big
opaque black shadow under the window, so clip out the region under
the terminal from the shadow we draw.

Add meta_window_get_frame_bounds() to get a cairo region for the
outer bounds of the frame of a window, and modify the frame handling
code to notice changes to the frame shape and discard a cached
region. meta_frames_apply_shapes() is refactored so we can extract
meta_frames_get_frame_bounds() from it.

https://bugzilla.gnome.org/show_bug.cgi?id=635268

NOTE: Applied only partially, compositor part is still missing...