Merge weights of independently trained models with no cost to loss

Big result for the dream of federated machine learning without prohibitive latency overhead