improve Dataset.create_torch_dataset and StandandLearnedReconstructor, enable more pickling

[jleuschn]

• fix CachedDataset for the case of all parts being cached. It supports random access then, but previously it was broken due to still calling self.dataset.get_sample[s] with the out parameter being all False.
• improve Dataset.create_torch_dataset
  • add transform parameter
  • move the previously local torch dataset class definitions to dival.util.torch_utility
  • use torch.utils.data.IterableDataset (instead of using torch.utils.data.Dataset and ignoring the index) for generator-based datasets
  • fix docs
• make 'ellipses' standard dataset pickleable and add fixed_noise_seeds parameter (still True by default, in contrast to noise_seeds (used for the ground truth image generation), which is False by default. Both (different) defaults are kept like they were).
• fix and improve evaluation.run implementation by avoiding to save reconstructions when run with save_reconstructions=False, reducing RAM usage
• add example scripts
• add option to allow multiple workers with generator-based datasets in FBPUNetReconstructor, remove warning about shuffling not working for generator-based datasets (for shuffle=True, it will raise an error due to Dataset.create_torch_dataset now returning a torch.utils.data.IterableDataset, and for shuffle='auto' the docs for StandardLearnedReconstructor describe that shuffling is not used in this case)
• improve StandardLearnedReconstructor
  • add possibility for a transform (data augmentation) via implementing init_transform in a subclass
• add shuffle option, defaulting to 'auto' which depends on whether the dataset supports random access. Previously shuffle=True was always used, but for generator-based datasets the index was ignored, effectively disabling shuffling.
• add dival.util.odl_utility.patch_ray_trafo_for_pickling