Efficient Long-Decoding Inference with Reasoning-Aware Attention Sparsity