Extending the Context of Pretrained LLMs by Dropping Their Positional Embeddings