Feature request: Support splitting model weights and training states into separate checkpoint files

Nobody has claimed this yet.

### Description & Motivation

# 🚀 Feature Request: Support splitting model weights and training states into separate checkpoint files

## Feature Request

Currently, PyTorch Lightning saves the entire training state (model weights, optimizer states, scheduler states, trainer state, etc.) into a single `.ckpt` file.

I would like to have an option to **separate model weights (and config) from training states** when saving checkpoints.

For example, the desired checkpoint structure could look like this:

```
checkpoints/
pretrained_model/
config.json # model configuration
model.safetensors # model weights only
training_states.pth # optimizer, LR scheduler, trainer states
```

### Pitch

_No response_

### Alternatives

_No response_

### Additional context

_No response_

cc @lantiga @borda

Start by locating PyTorch Lightning’s existing checkpoint save and load entry points, since no files or tests are named. Compare the current single .ckpt structure with the requested config.json, model.safetensors, and training_states.pth layout. Done should include a defined option for separate files and a clear way to restore training state.

Written by the indexing model from the issue text.