Efficient Many-Shot In-Context Learning with Dynamic Block-Sparse Attention