Improved Bounds for Pure Private Agnostic Learning: Item-Level and User-Level Privacy