Instead of instantly invalidating, we now cache the old render node and
do the update in an idle handler.
While that gives us a 1 frame delay, it avoids all the tricky things
like queueing resizes while resizing or queueing draws while drawing.
The only remaining issue (and a *big* one at that) is that a nested
widget paintable will now cause the widget to snapshot its previous
render node when creating a new one. And that one will snapshot its
previous render node, and that one will...
And nothing so far breaks this recursion.
This is actually not just a mechnaism to protect against too many
signals, but it's also a method to getting those signals at the wrong
time.
For every size/content change, a widget needs to invalidate twice:
Once when it queues a resize/redraw (going valid => invalid) and once
when the new size/content is actually assigned (going invalid => valid).
However, one of those invalidations might be inconvenient for the
listener. GtkImage for example does not like receiving
invalidate-contents signals when new contents are assigned, but is fine
with them when the old ones go invalid. And it will not try to draw the
paintable in between anyway.
So by bypassing the 2nd emission if nothing was changed, we can make
GtkImage happy.
