Stack Attention: Improving the Ability of Transformers to Model Hierarchical Patterns