How does this differ from XLA? Would tinygrad's lazy approach also just see the same unrolled loop right before compilation?