can't fit with ddp_notebook on a Vertex AI Workbench instance (CUDA initialized)

Nobody has claimed this yet.

### Bug description

Using this minimal code example:
```
import torch
import lightning as L

print(torch.cuda.is_initialized())
trainer = L.Trainer(
accelerator="auto",
strategy="ddp_notebook",
devices="auto",
max_epochs=1,
# callbacks=callbacks,
log_every_n_steps=1
)
print(torch.cuda.is_initialized())
```

On Google Colab with a T4 attached, both print statements print "False" as expected.

On a Vertex AI Workbench instance with a T4 attached, the second statement prints "True"; merely instantiating the Trainer initializes cuda. This prevents fitting with DDP.

What could be causing this, and is there any way to work around it?

### What version are you seeing the problem on?

v2.2

### How to reproduce the bug

_No response_

### Error messages and logs

```
# Error messages and logs here please
```

### Environment

Current environment

```
#- Lightning Component (e.g. Trainer, LightningModule, LightningApp, LightningWork, LightningFlow):
#- PyTorch Lightning Version (e.g., 1.5.0):
#- Lightning App Version (e.g., 0.5.2):
#- PyTorch Version (e.g., 2.0):
#- Python version (e.g., 3.9):
#- OS (e.g., Linux):
#- CUDA/cuDNN version:
#- GPU models and configuration:
#- How you installed Lightning(`conda`, `pip`, source):
#- Running environment of LightningApp (e.g. local, cloud):
```

### More info

_No response_

cc @justusschock @lantiga

Start with the minimal Trainer example in the issue and compare torch.cuda.is_initialized() before and after Trainer construction on Vertex AI Workbench versus Colab. Check the Vertex environment details and DDP notebook initialization path; done means identifying why Trainer initializes CUDA there and documenting a verified workaround or fix.

Written by the indexing model from the issue text.