Think-at-Hard: Selective Latent Iterations to Improve Reasoning Language Models