From Universal Language Model to Downstream Task: Improving RoBERTa-Based Vietnamese Hate Speech Detection