Remove output vocabs

[tdozat]

Currently, you can change the configuration settings of a model to let it output different things--for example, if you want to train a joint tagger/parser model, you could set the output_vocabs to tags:heads:rels.

But, this doesn't actually work. You have to hardcode in what the output of an nlp model is in a couple different places, so this configuration option is misleading and should be discarded. In the one or two places where it's used, its contents should be inferred automatically.