Optimizing Native Sparse Attention with Latent Attention and Local Global Alternating Strategies