Solves #131

Open Questions:

• Do we want special handling of lazy_tensors (data augmentation). This would require annotating the preprocessing operators with whether they are stochastic or not. If a lazy_tensor is stochastic, we should not allow using the tensor_dataset parameter

  Answer: I think it is better to implement something like po("materialize_lazy_tensor"), which takes in a Task with one or more lazy tensors and converts them internally to a tensor_dataset, which will make the dataloading efficient.

TODOs:

• allow to select which ingress operators to convert into a tensor_dataset