Learning Cross-scale Correspondence and Patch-based Synthesis for Reference-based Super-Resolution