SLA, combination of full attention + linear attention to sparsify attention in diffusion transformers https://www.arxiv.org/pdf/2509.24006