Refactor/SK-1144 | Simplify on_train and on_validate

[benjaminastrand]

Key changes

• Moved boiler plate code from client to ClientAPI to make it simpler for the user to implement an on_train/on_validate function
• Updated client_v2 accordingly
• These changes eliminate the client's (client_v2 or exemple in README) need to import fedn protos

grpc_handler.py

• Removed unnecessary params from send_model_update()

client_api.py

• Replaced _subscribers with train_callback and validate_callback
  • train_callback needs to take in_model as parameter and return out_model and metadata dict
  • validate_callback needs to take in_model as parameter and return metrics dict
• _task_stream_callback now invokes update_local_model or validate_global_model depending on task request type
  • update_local_model and validate_global_model contain boiler plate code that was previously inside the on_train and on_validate callbacks (getting/sending model from/to combiner, generating uuid, assembling metadata, sending model update/validation)
  • This replaces the functions train and validate
• run method creates a child thread for heart beats and listens to task stream in main thread

client_v2.py

• Adjusted training and validation functions for the updated client_api.py

README.rst

• Updated example client that makes use of the new API

[niklastheman]

Good improvements overall, nice work! I have some input (nothing major), check out the comments!

[niklastheman]

I think we could include a callback function for training in this README example. It doesn't need to do anything more than return the input model. The end goal is to make users understand how they can develop clients that help with their use case.

[niklastheman]

Not sure about these changes... Don't know if a TaskRequest as input is the way to go. It creates a dependency that did not exist previously... If you think it's cleaner, keep it.