The Devil is in the Details: On the Pitfalls of Vocabulary Selection in Neural Machine Translation