Robustly overfitting latents for flexible neural image compression