Boosting Cross-Domain Speech Recognition with Self-Supervision