Lightning 2.0 CPUAccelerator is extremely slow!

Nobody has claimed this yet.

### Bug description

I just updated lightning version to 2.0 and running inference on cpu is extremely slow.

My previous code can be [found here](https://github.com/Unbabel/COMET/blob/master/comet/models/base.py#L616). Inference is basically performed by initialising a trainer with devices=0 and strategy=None.

```python
trainer = ptl.Trainer(
devices=devices,
logger=False,
callbacks=callbacks,
accelerator=accelerator if gpus > 0 else "cpu",
strategy=None if gpus < 2 else "ddp",
enable_progress_bar=enable_progress_bar,
)
return_predictions = False if gpus > 1 else True
predictions = trainer.predict(
self, dataloaders=dataloader, return_predictions=return_predictions
)
```

Since in lightning 2.0 we cant pass `None` to the strategy, I replaced it with "cpu" if gpus < 1. Yet, this makes the code much slower. Is this normal? In gpu everything seems to work well.

### 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 @borda

Start with the Trainer.predict call shown in the issue and compare CPU inference using the Lightning 2.0 strategy settings with the previous configuration. Collect the missing Lightning, PyTorch, Python, OS, and installation details, then create a reproducible benchmark. Done means the CPU slowdown is reproduced and its cause or a documented resolution is established.

Written by the indexing model from the issue text.