A Quantitative Approach to Understand Self-Supervised Models as Cross-lingual Feature Extractors