On the Limitations of Steering in Language Model Alignment