Accelerator has an example experiment that uses Secure Aggregation

[thomasp-ms]

Is your feature request related to a problem? Please describe.
Currently, if customers wants to add Secure Aggregation (SA) to their Federated Learning (FL) experiment, they are on their own.

Describe the solution you'd like
We would like a solution "a la shrike", where all the customers have to do is implement a few key functions (named generate_noise() and anonymize_model_weights() in shrike). Then the factory code will appropriately handle applying SA.

Describe alternatives you've considered
Haven't considered alternatives, but open to discussion.

Additional context

• Here is the documentation for the shrike FL API. Check it out for instructions on how things are done with shrike.
• Here is the shrike code for the FL API. Look for the generate_noise() or anonymize_model_weights() functions and how they are used, or look for use_secure_aggregation to locate the SA-related parts.
• The pre-built components to generate noise and anonymize model weights can be found here and there.
• An example pipeline using the 2 components above can be found here, and there is the associated config file.

This issue is about adding an example experiment, consuming the code and components introduced in issues #90 and #91.