CFDRM: Coarse-to-Fine Dynamic Refinement Model for Weakly Supervised Moving Vehicle Detection in Satellite Videos