Adding the "train" argument back to "postprocess"

[jwohlwend]

I'd like to make a case for adding the "train" argument back to the Field's postprocess call. Fields tend to be shared across training and test sets, so the train argument was a flexible way to apply different postprocessing at training and test time. One example: using "postprocessing" to add noise to the input text dynamically. Without the train argument, I don't think there is a clean way to deactivate the noise at test time. Happy to take suggestions for work arounds, but I think having the train argument should generalize to more use cases.