Large Language Model as Attributed Training Data Generator: A Tale of Diversity and Bias