LSG Attention: Extrapolation of pretrained Transformers to long sequences