Add MNIST / EMNIST to benchmark datasets

[Saipraneet]

It is probably a good idea to include a "standard" synthetic baseline (say EMNIST or MNIST). We can use the performance on these well known datasets to convince the readers/reviewers that our strategy implementation + hyperparameter tuning is correct. This should be pretty quick to do given the extensibility of FLamby.