Cross-validation while still enabling CLI control

Nobody has claimed this yet.

### Description & Motivation

Related issue:
https://github.com/Lightning-AI/pytorch-lightning/issues/20544

Hi Lightning team,

I saw the discussion in the issue above about supporting cross-validation in LightningCLI. I tried implementing a simple approach where the CLI is used as a reusable training template, and cross-validation is orchestrated outside the CLI.

The idea is to wrap `LightningCLI` inside a function (e.g. `cli_main`) and call it multiple times with different fold indices passed through CLI arguments.

For example:

```python
def cli_main(args):
cli = LightningCLI(
model_class=MyLightningModule,
datamodule_class=KFoldDataModule,
args=args,
run=False,
)
cli.trainer.fit(cli.model, cli.datamodule)

def arg_parse():
pass

other_args, lightning_args = arg_parse()

# cross-validation orchestration
for k in range(5):
cli_main(
lightning_args + [
f"--data.k={k}",
f"--trainer.logger.init_args.version=cv-fold{k}",
]
)
```

In this setup LightningCLI acts as a reusable training template and all configuration is still controlled through CLI arguments.

Since this pattern works quite naturally with the current CLI design and doesn’t require much extra code, I was wondering whether something along these lines could potentially be integrated into LightningCLI in the future.

Curious to hear your thoughts.

### Pitch

_No response_

### Alternatives

_No response_

### Additional context

_No response_

cc @lantiga @mauvilsa

Start by reviewing related issue #20544 and the LightningCLI usage shown in this issue. Evaluate the proposed reusable cli_main pattern, repeated fold arguments, and logger versioning, then define what an integrated cross-validation interface should do and how completion would be validated.

Written by the indexing model from the issue text.