Add text-conditioned model (CLIP cross-attention, 120k steps)

sprite-gpt: rectified-flow diffusion for emoji + sprites (32px uncond, 64px class-cond)

initial commit