This also changes the way we insert nodes into a
MultiChildRenderObjectWrapper's renderObject, which fixes issue #626.
Now, instead of the slot being a renderObject, it's the Widget that
currently uses that renderObject. That way when the Widget changes
which renderObject to use, the siblings of that Widget in the same
child list don't have to be notified of the change.

I tested performance of the new algorithm vs the old algorithm using
the stocks demo at idle and the stocks demo scrolling steadily. The
data suggests the algorithms are roughly equivalent in performance.