Implement full set of benchmark datasets evaluated by Andrew et. al.

[lguerdan]

Currently, we have implemented CIFAR-10 and MNIST datasets, which were used by Abadi et. al. in the DP-SGD paper. Andrew et al (https://arxiv.org/abs/1905.03871) evaluate a wider set of models on CIFAR-100, EMNIST CR, EMNIST AE, Shakespeare, StackOverflow NWP, and StackOverflow LR.

Add an additional set of benchmark datasets, ideally including language datasets such as Shakespeare and StackOverflow NWP. Torchvision includes an additional set of useful datasets that might be easy to add in our current code API.

Adding a dataset involves:

• Creating the relevant backbone model in the models/ folder
• Adding the model loader to the load_model function in train.py
• Adding the dataset option in the load_data() function in utils.py
• Adding the benchmark in the run_exp() function in train.py

For some more complex datasets, it might require adding additional pre-processing code to load dataset batches in the correct form.