Because if you're an actor with a lot of faces (eg, state level actor with a face db) to feed it, and you're sweeping large numbers of phones rather than trying to hit one in particular, this suggests they'd have an easier time than it sounds like when you're considering it in the private use context.
I think it can be assumed that the facial recognition is rate-limited, just like PIN entries. Even if you had a million phones and a million faces in your database, you could only try perhaps 30 faces on each phone.
If you have the kind of surveillance tech people are postulating here, you already have a high-res-enough scan of the actual face of the person you target in order to produce something to unlock their phone.
Not likely - they used professional hollywood mask makers to test against... Remember, this is infrared with a 30k dot projector - the most accurate 3d visual record you could make would appear to be insufficient.