Yeah I think it's best to wait for accelerated surface before introducing them, and then only do it on a as-needed basis.
Done, including some extra logic to be able to properly handle reparenting, as we move between child/non-child states.
I don't know how this will work in Wayland, but on X11 we ended up (not upstream but in Proton) with multiple client windows for a single HWND, because of how Windows surfaces can work.
In the Wayland driver the plan is to create a separate `wl_subsurface` (making it a child of the main `wl_surface` for the HWND) to hold any accelerated rendering buffers. This will apply to both child window accelerated rendering and main window accelerated rendering.