You _can_ use various hooks to dynamically bind and unbind on a single GPU system but that would be another degree of difficulty for newcomers (on top of potential buginess throughout the stack).
As to the latter part of your first sentence, I believe that virtio-gl works in the manner you describe. Thought it is Linux only. I have experimented with it on Proxmox and it works well enough.
VFIO does work, but it's hacky, so it's very inconvenient unless one has plenty of time and patience.
For example, AMD GPUs don't support rebar (took a while to figure out), and they used to have (I don't know if it's been definitely solved) a problem with reset, which caused random lockups when exiting the VM.
Additionally, power saving/device ownership must be handled. Which means: either one terminates the X session each time before starting/after stopping the VM (so X will handle the power management), or they leave the card owned by the vfio driver (but this makes the card run hot while sitting doing nothing, because the vfio driver doesn't handle the power management), or they swap drivers on the fly hoping that it will work fine (on nvidia, it's not a stable strategy; no idea on AMD).
And these are just a few issues. I've experienced others in the past, and it's very frustrating and time consuming.
With all these problems, a combination of dual boot (for demanding 3d tasks) and VMWare (for light ones) is much simpler, and stable. I hope to replace VMWare with VirtualBox for the latter use case.
Hard to tell if that's what's advertised in these release notes, but if so, then at least older games might finally exhibit some decent performance.