Save initial and latest models in addition to best

[sarthakpati]

Is your feature request related to a problem? Please describe.
In a lot of cases, the latest model might be more informative WRT how the training loss has been behaving as compared to the latest. Also, the initial model can give greater insight into post-training data.

Describe the solution you'd like
In addition to the best model, save initial (i.e., the model right before any training has occurred) and latest models.

Describe alternatives you've considered
N.A.

Additional context
N.A.