Kml as training set for S2 classification: serious problem

to have points as training samples they need to have the followings attribute structure

ID, class
1, forest
2, forest
3, water
4, urban
5, forest
6, urban

If you manage to convert the attributes of your KML imported into QGIS into this structure, SNAP will recognize it correctly.

As collecting a sufficient number of training points can be exhausting, maybe these considerations help you as well: Number of training samples at Random forest classifier