CoSeR: Bridging Image and Language for Cognitive Super-Resolution