Refer, Reuse, Reduce: Generating Subsequent References in Visual and Conversational Contexts