Pre-training Contextualized World Models with In-the-wild Videos for Reinforcement Learning