It's because if the GPU is reverse engineered enough, you can simulate an Apple GPU on another machine _and_ run other OSes with a proper GPU stack on those macs.
Most vendors of graphics hardware keep their interface secret, but not all. The problem is that the set of hardware that isn't secret won't overlap much with what people have or want to run, which is pretty central to what makes Hackintoshes attractive.
Now I don’t know anything about GPU stacks, but I could imagine that one change essentially requiring a rewrite of the entire stack if certain assumptions were baked into the old one.
The new model would seem to allow the developer to allocate a buffer full of stuff and rapidly alternate between CPU and GPU instructions on that buffer, without copying. If so, that’s a pretty huge win for performance.
Currently, Apple and NVIDIA have the beefiest UMA w/ the CPU GPUs available, and those are Arm solutions.