Learning Hybrid-Control Policies for High-Precision In-Contact Manipulation Under Uncertainty