fixed federated learning example

[wilko77]

For gradient decent in a federated setting to work, it is essential, that all parties start with the same model.
In our example we first let the clients learn a model on their respective data, and then we ran federated learning on top of these individually learned, different models. So sad.

Here I propose to better separate the local learning from the federated learning part. In both cases, all clients will start with an initial model of zeros.

The output now is:

Loading data
Error (MSE) that each client gets on test set by training only on own local data:
Hospital 1:	3921.78
Hospital 2:	3808.58
Hospital 3:	4019.43
Running distributed gradient aggregation for 50 iterations
Error (MSE) that each client gets after running the protocol:
Hospital 1:	3644.47
Hospital 2:	3644.47
Hospital 3:	3644.47