Feature Request - lifecycle callbacks for SupervisedTraining

For SupervisedTraining/Others?, it would be useful to add the following optional lifecycle methods:

• per-epoch callback (lambda/function params/context to be discussed, but minimally the epoch number for context)
• per-batch callback (lambda/function params/context TBD, but perhaps a good start would be the actual batch samples, targets and the epoch number for context)
• other points of interest for the training loop, with related context?

In the meantime, a custom training loop works fine, but this might be helpful for others.