I can't see what benefit you expect to get from the GPU. It is plenty fast with all-CPU rendering.

The Qubes X model is interesting: each VM runs its own X, and renders windows into memory shared with a central VM that copies the pixels to real video memory. Input events get forwarded to a VM's X according to which is the active window frame. The camera can be attached temporarily to a VM, and mike input similarly. So, no VM has physical access to hardware anytime other than when you want it to.

I am hoping a future Qubes will have proxying Vulkan stubs or something, for controlled access to a GPU. But it isn't missed much since FF stopped crashing.