Time Matters: Enhancing Pre-trained News Recommendation Models with Robust User Dwell Time Injection