Knowing the Distance: Understanding the Gap Between Synthetic and Real Data For Face Parsing