[Sparse Attention] Native Sparse Attention (NSA) Explained: Efficient Long-Context Modeling for LLMs