MPC Based Secure Aggregation for Federated Learning

[iamtrask]

This issue depends on #1501 being completed.

This issue requires creating a demo using MPC for the secure aggregation of gradients. Multiple different workers should create a weight update (in the usual Federated Learning fashion) but instead of sending the gradients directly to the model owner, they should use MPC to securely aggregate their gradients wherein only the aggregated update is sent to the model owner.

[iamtrask]

Issue #1501 has been completed - this issue is ready for someone to pick up. I'm dealing with some broken unit tests but i'll pick it up when i'm done if no-one else has.

[iamtrask]

I'm picking this up now.