I have looked into this extensively, and all I can say is that it's extremely difficult to get working on certain hardware configurations if you only want to use one card, and even then it's hard to set everything up the way you want to do it.
I've been running 2 Ubuntu servers running 3 gaming VMs. I agree to a certain extent that hardware choice is key. I got burned by AMD GPUs not resetting so switched to NVidia GPUs.
Didn't have issues for almost a year.
Rule of the thumb for me is: AMD GPU for host, NVidia for guests. I use AMD GPUs for host because open source drivers are awesome and makes kernel/version updating easy. NVidia for guests is no brainer since they don't have PCIe reset issues. For motherboard, pick one that has good IOMMU groupings, usually x<3|4|5>70 for AM4 or x299 for Intel, or go Threadripper.
There's actually a kernel patch that basically separates every pcie device into its own iommu group (linux-vfio on arch). You shouldn't do that if you need the security provided by proper iommu groups, but it works for a gaming vm.
According to Wendell on Level1Linux, the reset bug should be fixed on AMD 6000 series gpus.
With tools like Looking Glass (https://looking-glass.io/) you don't even have to leave the Ubuntu Desktop environment.