PCI-passthrough/GPU support

[alexkreidler]

This would mainly be useful for machine learning development in Docker, and would allow us to use https://github.com/NVIDIA/nvidia-docker (or similar) to build and run GPU containers.

See https://github.com/mist64/xhyve/issues/108.

[benaubin]

Would be great to use with an @NVIDIA GPU!

[rn]

As pointed out in the link to xhyve above it is non-trivial and not at all clear if Apple exposes the right APIs for this (e.g. access to the IOMMU from userspace). GPUs are further complicated as they are obviously also used the host. So GPUs would require either on device virtualisation capabilities (so that a virtual GPU could be passed to the VM, assuming VT-d is accessible) or some higher level GPU proxy device.