Loading checkpoint from CLI using SLURM doesn't use GPU even though it says it does

Nobody has claimed this yet.

### Bug description

When I load my checkpoint, it says LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]

But I check my slurm jobstats and it's like

GPU utilization per node
stellar-m01g3 (GPU 0): 0% <--- GPU was not used

GPU memory usage per node - maximum used/total
stellar-m01g3 (GPU 0): 12.2GB/40.0GB (30.5%)

I even made sure to later put accelerator: gpu in the trainer section of the yaml file

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

v2.5

### How to reproduce the bug

```python
cli = ModelCLI(
subclass_mode_model=True,
subclass_mode_data=True,
parser_kwargs={"parser_mode": "omegaconf"},
save_config_callback=None,
)

trainer:
max_epochs: 10
accelerator: gpu
enable_progress_bar: False

#SBATCH --gres=gpu:1
#SBATCH --time=04:00:00

module purge
source ...

export SLURM_JOB_ID=$SLURM_JOB_ID

srun python -m specseg.models.train \
fit \
--config config/model/config_label.yaml \
--ckpt_path /path
```

### Error messages and logs

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

### Environment

Current environment

```
#- PyTorch Lightning Version (e.g., 2.5.0):
#- PyTorch Version (e.g., 2.5):
#- Python version (e.g., 3.12):
#- OS (e.g., Linux):
#- CUDA/cuDNN version:
#- GPU models and configuration:
#- How you installed Lightning(`conda`, `pip`, source):
```

### More info

_No response_

cc @lantiga

Start with the ModelCLI invocation in specseg.models.train and the provided config/model/config_label.yaml, then inspect how the trainer's accelerator setting is passed under the srun command. Reproduce with the v2.5 setup and compare the reported LOCAL_RANK/CUDA_VISIBLE_DEVICES values with SLURM GPU utilization; done means the checkpoint-loading CLI job demonstrably uses the allocated GPU.

Written by the indexing model from the issue text.