Bridging the Language Gaps in Large Language Models with Inference-Time Cross-Lingual Intervention