Revisiting semi-supervised learning in the era of foundation models