Saving the Latest best checkpoint

Nobody has claimed this yet.

### Description & Motivation

I am working with Graph Neural Networks and training models to learn graph structures and test their expressiveness. In such a case, acc=100% happens.

I adopted a randomized approach, where once in a while, due to variance, the validation score gets 100% high during the early stage of training, while the average validation ACC across epochs is low. The validation results get more stable and also reach 100% sometimes in the later stage of training, and I want to save these checkpoints, but the earlier 100% score prevents the ModelCheckpoint from saving them, and I can only save the latest checkpoint, which might not be optimal.

Is it interesting to add save_latest_best_checkpoint mode to ModelCheckpoint, so we break ties in favor of later epochs?

### Pitch

Add save_latest_best_checkpoint mode to ModelCheckpoint

### Alternatives

_No response_

### Additional context

_No response_

cc @borda @carmocca @awaelchli

Start by reading the ModelCheckpoint implementation and its validation-score comparison behavior. Determine how an equal best score is handled, then add coverage showing that the proposed mode retains the checkpoint from the later epoch; done means later ties are saved without changing existing behavior.

Written by the indexing model from the issue text.