Bridging the Gap between Language Models and Cross-Lingual Sequence Labeling