vGPU scheduler assume all the nodes have GPU information annotation. Cannot handle cpu node or the period before annotation got patched

[Fizzbb]

To reproduce, when a new node join, all the daemonSet on the node stuck at pending node, with PostFilter plugin error messages.
The is caused by filter plugin denied the node (no GPU annotation found)

[YHDING23]

When a new node joins the cluster, all the DaemonSets should be scheduled by our customized scheduler. The old scheduler design needs to capture some node information (including GPU counts) of this new node, then make the scheduling decision of these DaemonSets. However, without deploying the DaemonSet of the device plugin as a pod, no GPU information can be captured.

The issue is fixed by adding a function, IsSharingGPU, to determine whether a pod is sharing GPU or not. If the pod is not planning to use any GPU, it will be scheduled without using any node-level information.