well , it must be some errors
As far as I know, DxO tests just one copy of a lens. Given Sigma's reputed less-than-stellar QC (not than Canon/Nikon are perfect, far from it). So, perhaps not errors on DxO's part, but rather copy variability as seen in Roger Cicala's testing of his rental stock.
P-Mpix is actually a computed number (well, MTF is computed as well but P-Mpix is computed from the MTF data). How is it computed - it is a well kept secret. Until they explain what it is, it is worthless.
SQF is one of the many attempts to represent the whole MTF curve by one number. It depends on a reference viewing size. As such, it is not better than, say, MTF-50.
I wouldn't say worthless - unlike their Sensor Score, which introduces bias with the method, nothing in the description of P-Mpix suggests it's other than a linear transform of SQF measurements.
I'd say SQF is not necessarily better than MTF50 as a single-number comparator, but perhaps more relevant/useful, in that the values more closely approximate how a viewer would rate picture sharpness (which isn't always the case for MTF50).