Fix a bug where intermediate phase outputs could get too low height.

Basically our aspect ratio adjustment and lack of proper roundoff
could conspire to let e.g. mix(1280x720, 939x939) = 1669x938,
which is obviously wrong. I could probably have fixed it with
an extra lrintf(), but it seems more robust to simply avoid the
extra conversion (ie., never convert height -> width -> height).

Add a unit test to verify.