My suspicion is that they are not stored at all, but I thought asking couldn’t hurt 
Listens in the data dumps are very valuable, but I think if more user behaviors (e.g. start, end points of listens, whether a song is skipped, etc.) are logged, it can help train better recommendation systems
We don’t collect that data mostly because we don’t have access to it. Unlike the streaming services we can’t really track each click that users make. The best we can do is record the listens and we could attempt to infer skips (e.g. the next track followed sooner than the play time of the track) but even that is fraught with peril.
Not to mention possible privacy issues.