For the Intel drivers, using glReadPixels() to read into client-memory
directly from the frame buffer is much slower than creating a pixel
buffer, copying into that, and then mapping that for reading. On other
drivers, the two approaches are likely to be similar in speed. Create
a ShellScreenGrabber abstraction that uses pixel buffers if available.
Use that for screenshots and screen recording.
https://bugzilla.gnome.org/show_bug.cgi?id=669065
