Composite Active Learning: Towards Multi-Domain Active Learning with Theoretical Guarantees