Contact Optimization with Learning from Demonstration: Application in Long-term Non-prehensile Planar Manipulation