Dynamic Context Pruning for Efficient and Interpretable Autoregressive Transformers