Cross-domain Random Pre-training with Prototypes for Reinforcement Learning