That is quite odd.
Don’t tell me software emulated GPU can’t be less attack surface than hardware GPU, the fact this is not supported even for iGPUs is truly beyond my comprehension.
Is there not enough developers to work on such features? I am willing to volunteer.
It is less surface while at the same time reducing priviledged access, isolating firmware from user data etc.
(See chapter 5 of the referenced Qubes OS acrhitecture pdf for basic considerations at the beginning of the project. It explains a lot, though outdated.)