Fix dynamic accessibility tree updates not working in the interpreter

[tronical]

It turns out that for repeated elements, we set root_weak to self_weak, instead of the root we obtain from the parent instance. Fix this by moving the code to initialize parent_item_tree_offset to before we try to use it (via parent_instance) to initialize root_weak.

[ogoffart]

This is interresting that only accessibility broke. Is nothing else using the root_weak?

[tronical]

I suppose like many other places, we just silently skip. Like in the accessibility case, we just didn't end up calling unregister_item_tree because of self.root_weak().upgrade()?;

[tronical]

But indeed, there are only three call sites for maybe_window_adapter. One in instantiate (that always works), one in Drop (this one didn't), and one in the ffi, called from the C++ interpreter API I suppose.