Node Auto-Provisioning failing for certain GPU nodes (T4)

## How to re-create
A job that is marked as requiring `nvidia.com/gpu`, if results in a new node being spun up in GKE, will _fail to be scheduled_ on that node.

## Why is this bad
- Using GPU nodes with Node-Auto-Provisioning in GKE is **broken** (at least for T4s, not sure which other GPU types are affected)
- It feels strange that such a core "elasticity behavior" is unacknowledged -- hoping this issue gets attention and results in at least an ETA for the fix

## Details on error

The provisioned node has a `nvidia-device-plugin` pod
This pod has a `nvidia-driver-installer` container which is an `init` container
This container is stuck on startup

```
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
0 0 0 0 0 0 0 0 --:--:-- --:--:-- --:--:-- 0 100 720 100 720 0 0 113k 0 --:--:-- --:--:-- --:--:-- 117k
GPU driver auto installation is disabled.
Waiting for GPU driver libraries to be available.
```

As a result, the kubelet never registers the `nvidia.com/gpu` resource, which means that the job (which triggered the node in the first place!) can't get its pods scheduled on it.

## Prior context:
This is based off the following issue, which is _no longer fixed_ (but which I cannot reopen)

https://github.com/GoogleCloudPlatform/container-engine-accelerators/issues/356