AttackerModel should not perform normalization

[NoahAmsel]

I think it would be cleaner if the normalization step were baked into the underlying model rather than being performed by the AttackerModel. Conceptually, the act of normalizing inputs is not related to the AttackerModel. Combining them causes problems. For instance, I'm trying to wrap AttackerModel around a network that already does normalization. My proposed change is to make the dataset optional in the initialization of an AttackerModel:

    def __init__(self, model, dataset=None):
        super(AttackerModel, self).__init__()
        if dataset is None:
            # we can skip the normalize step if we want
            self.normalizer = ch.nn.Identity()
        else:
            self.normalizer = helpers.InputNormalize(dataset.mean, dataset.std)
        self.model = model
        self.attacker = Attacker(model, dataset)

[andrewilyas]

This is a good idea! Am a little swamped right now, but will get it into the next release. Pull requests into the develop branch also welcome :)