Get rid of `to(torch.[type])` that litters the code.

[evanberkowitz]

We should probably just leverage something like torch.set_default_tensor_type(torch.DoubleTensor) at the beginning of scripts instead.