StEP: Style-based Encoder Pre-training for Multi-modal Image Synthesis