Rethinking Word-Level Auto-Completion in Computer-Aided Translation