Normalize both POT files before comparing.

Otherwise we'll often converge to a state where we always get into the
"Real changes ..." case, which feels wrong; after a single run,
consecutive runs should always end up in the "Only header ...# case.