Evaluating the Impact of Text De-Identification on Downstream NLP Tasks