Learning How to Forget: Fine-Tuning for Long-Context Sparse Attention