Assessing the Latent Automated Program Repair Capabilities of Large Language Models using Round-Trip Translation