Thinking Outside of the Differential Privacy Box: A Case Study in Text Privatization with Language Model Prompting