Support running on nodes with host-installed GPU drivers

[zerodayyy]

Nos is currently broken on systems where GPU drivers are pre-installed on hosts, for example on AKS. The symptom is gpu-agent Pod not starting due to missing /run/nvidia path on host.

According to Nvidia DRA driver documentation, /run/nvidia folder is provided via driver container. When drivers are installed on host instead of via container, the path is missing and has to be symlinked to host root manually:

Ensure your NVIDIA driver installation is rooted at /run/nvidia/driver

For deployments running a driver container this is a noop.
The driver container should already mount the driver installation at /run/nvidia/driver.

For deployments running with a host-installed driver, the following is sufficient to meet this requirement:

mkdir -p /run/nvidia
sudo ln -s / /run/nvidia/driver

NOTE: This is only currently necessary due to a limitation of how our CDI
generation library works. This restriction will be removed very soon.

To implement support for host-installed drivers, we can simply mount host's / as /run/nvidia/driver inside gpu-agent container.