Self-Supervised Learning of Visual Servoing for Low-Rigidity Robots Considering Temporal Body Changes