Fix prediction of "negative-n" sources

Sources that are in the direction of negative w should get a negative n value, but Dp3 treats them as if they are positive w. This MR fixes that. Sources for which this holds are more than 90 degree away from the phase center. This will not occur so often in field-based self-cal, but will arise quite commonly in a-team subtraction or demixing. It is thus somewhat surprising that this bug has existed for so long...

I did a quick check to see if negative n-values can pose problems elsewhere, but I think this is not the case. I'll ask Satyapan (who found this issue) to run his tests again to check negative-n sources work after this.