Cross-modal Domain Adaptation for Cost-Efficient Visual Reinforcement Learning