What does the analysis of the Upwork case study suggest about why the gap in likelihood of hiring between recommended and not recommended disappeared after the interview stage?