Flamingo: a Visual Language Model for Few-Shot Learning