Fixes issue #482.

The web script's presentObject() calls getObjectContentsAtOffset() to determine
what to present for that object. Normally that results in a list of one or two
content items to present. But when the object to be presented is the document
itself, the resulting content items will likely be in the hundreds (at least).
It's like doing a SayAll, but gathering the full content prior to saying anything.
This makes it appear that Orca has hung.

Thanks to the collection interface, it is possible to very quickly
look for any widgets that suggest a dialog is not a simple message
dialog. Unfortunately, in some apps (e.g. Pidgin, a Gtk2 app) this
seems to make Pidgin crash for reasons still to be determined. In
order to try to avoid this sort of crash, use several collection-
based checks to quickly locate the first container which suggests
that a dialog is not a message box. Hopefully this will prevent
AT-SPI2 from digging so deeply in the tree that Pidgin crashes,
and it's a reasonable change to do anyway.

See issue #485.

Apparently we cannot always get reliable values for the active
window position from Gdk. Therefore we need to modify our hack
in _window_coordinates_to_screen_coordinates.

1. Use AT-SPI to get the screen coordinates. This won't work in Wayland
   or with Gtk4 apps, but seems to work for Gtk3.
2. Because the window coordinates we get appear to not take the title
   bar height into account, calculate that by subtracting the reported
   height of the active Gdk window from the frame extents' height.