Draft: add a global seed for reproduceability

The idea here is to add a global seed policy, akin to the device policy, to ensure reproduceability of experiments. I see two steps :

• Wiring all existing seed mechanisms to the a get_seed_policy method, which include the seed for datasets.split_data, datasets.InMemoryDataset, and the optimizer.modules.NoiseModule.
• Creating new seeding mechanisms in model and training, notably for Dataset.generate_batches and in the different framework implementaion of model initialization and randomized layers (e.g. dropout).