Heterogeneous treatment effect bounds under sample selection with an application to the effects of social media on political polarization