Colby Cosh finds out what subsets, modelling assumptions and ‘non-probability samples’ have to do with polling these days