The offset (zero error) at 25°C is maximum +/-15°C. There is no stated error for the sensitivity (span error). So at 25°C die temperature it will likely read between +/-4000 counts or so.
So if you do a single-point calibration at room temperature there's no guarantee about errors at other temperatures.
It's not guaranteed as good as a dedicated temperature sensor because they've not spent the money in design and probably test fixture test time to make it good. It's a freebie, probably with some kind of digital calibration that takes place almost instantly. It may, in fact be much better than guaranteed, but they're definitely not promising anything of the kind.
As a general rule, if you read a datasheet and something seems ambiguous or suspiciously unstated, assume the worst and you'll very seldom be wrong.
If you assume the main purpose of the die temperature sensor is to compensate for temperature drift in the other sensors by calibration, the absolute accuracy is not very important, only stability (and resolution). Even linearity is not that critical.