Description
In some rare cases, for example, when you need to finetune a large model on a small dataset the majoring part of training loop is waiting for saving model checkpoints to a hard drive.

Proposal
Would be logically to add a CheckpointCallback with parameter save_n_best=0 to a configuration and do not store best checkpoints and instead use the latest state of the model.

Note
All of the described above is a proposal mostly for config API because config API during the stages loads the best checkpoint from the previous stage.

So here are a few steps to how it can be achieved:

1. For using the latest state of the model you need to overload a get_model property of ConfigExperiment to return the latest model (something like in BaseExperiment).
2. To prevent saving the model state to a hard drive overload property get_callbacks of experiment class and remove CheckpointCallback from default callbacks.
3. Create and use an empty callback instead of CheckpointCallback.