Distilled Dual-Encoder Model for Vision-Language Understanding