The only way to avoid the composition latency is to have the entire composition be done when the GPU is sending the final image to the monitor (note that performing composition using the GPU via OpenGL or whatever is not the same thing even if both are done using the GPU), pretty much like what "hardware mouse cursor" gives you. This would require GPUs to support an arbitrary number of transformable overlapping surfaces (where a single surface=a single toplevel window) and applications being able to draw to these surfaces directly without any form of intermediate buffering (which is important to avoid the initial late frame latency).

Now, it isn't like it is impossible to make a GPU with this sort of functionality since GPUs already do some form of composition already, but AFAIK there isn't any GPU currently on the market that can do all the above. At best you get a few hardcoded planes so you can implement overlays for fullscreen content.

And of course none of the above mean that you have to use Wayland, the X server could perform toplevel window composition itself just fine.