Mount host NVIDIA libraries such as NVML instead of packaging them in the image

[empovit]

The controller requires NVIDIA libraries to work. Today we have them inside the image, inherited from the CUDA base image. We should instead make the operator mount all required libraries from the host.

Important: This requires the libraries to be present on the host. On OpenShift it means having the NVIDIA GPU Operator deployed and a ClusterPolicy created.

As part of these changes, we can replace the base image with UBI.

[harche]

/cc @asm582 @sairameshv @mamy-CS @mrunalp