Allow checkpointing initial model weights

### Description & Motivation

Sometimes it can be useful to have access to the randomly initialized weights of a model prior to training. For instance, when evaluating against downstream tasks the untrained model can serve as a random baseline for comparison, especially since many publications are showing that the architectural prior of a randomly initialized model exhibits representational structures that can transfer to downstream applications without any training.

The current `ModelCheckpoint` implementation does not offer an option to allow saving the initial state of a model before any optimization is performed.

### Pitch

The `ModelCheckpoint` callback could be extended by a `save_initial_weights` option that would control whether the initial model is checkpointed prior to training. The flag would default to `False` to prevent breaking behaviour.

### Alternatives

* **Random state can be fixed.** However, this influences other components than the model initialization as well and prevents reproducing a certain model state by simply loading a certain checkpoint.
* **Custom subclass of `ModelCheckpoint`.** Works of course, but would be nice to see this feature as part of PyTorch Lightning.

### Additional context

_No response_

cc @borda @awaelchli @carmocca

Start by locating the ModelCheckpoint callback and tracing when it first saves during training. Add a save_initial_weights option that defaults to False and controls whether the model's pre-training state is checkpointed. Done means the default behavior is unchanged and enabling the option produces an initial checkpoint before optimization.

Written by the indexing model from the issue text.