rbtree: Move to an approach where we don't move contents

rbtree: Move to an approach where we don't move contents

So instead of copying the children and height to the new node, we keep
the old node and copy all the old stuff to it.

This is necessary so the accessibility code can use the node as a key in
the hash table or store the node as a reference to the row instead of
GtkTreeRowReference. And because it already does that (oops), this fixes
a bunch of segfaults with a11y enabled.