You Are What You Train: Effects of Data Composition on Training Context-aware Machine Translation Models