Return all available directly if preferred allocating all

[zwpaper]

We found retrieving all GPU statuses will take seconds in nodes with multiple GPUs when kubelet calls GetPreferredAllocation.

and if the available ones are all the kubelet requesting, maybe the device plugins can return the available ones directly.

we made some changes and found it was working as expected much faster.

so we are raising up this PR to discuss this solution:

1. is it ok to skip the following nvml detections in GetPreferredAllocation?
2. are there any other things we need to consider before adding this change?

[elezar]

Note that this repository is a read-only mirror of https://gitlab.com/nvidia/kubernetes/device-plugin.

If possible, please create a merge request there instead and close this PR.

[zwpaper]

Hi @elezar, thanks for the info, created one at https://gitlab.com/nvidia/kubernetes/device-plugin/-/merge_requests/339

[elezar]

@zwpaper sorry for the back and forth. We are in the process of migrating to GitHub as our primary repository. I am reopening this PR and closing the GitLab one.