Add initial_epoch to Segmentation.fit()

[hvgazula]

What would you like changed/added and why?

• I want the ability to resume training from a previous training run. See [link](https://www.tensorflow.org/api_docs/python/tf/keras/Model#fit:~:text=compile()%20instead.-,initial_epoch,-Integer.%20Epoch%20at)

What would be the benefit?

• see above

Does the change make something easier to use?

• see above

[hvgazula]

works well with init_from_checkpoint (might as well make this the defacto step after initializing a model)

[hvgazula]

feature: maybe cascade the number of completed epochs from here to initial_epoch in fit?