Use Epochs to load CIFAR10

[sgugger]

This PR starts the work of removing Batcher and use Epochs instead.

To avoid one gigantic PR, I have created new protocol named ImageClassificationData that I use and kept the old cifar model in OldCifar10 (to use it in the benchmarks). The protocol ImageClassificationData can be renamed once everything is done and the current ImageClassificationDataset is removed.
The OldCifar10 file can be removed when MNIST and the benchmarks have been converted to use Epochs.

New with this implementation:

• the shuffle of the training set can be seeded;
• the pixel values of the tensors are always scaled to [0, 1] even if normalize is false;
• the tensors are created on the fly from the raw UInt8 data to avoid launching several kernels of Tf-runtime

[BradLarson]

I think this makes sense to transition parts over. Once MNIST is done, we should be able to migrate the benchmarks across and see how this impacts performance.