Attention kernels that reduce memory/time complexity using tiling, FlashAttention, or linearized variants.