ROSO: Improving Robotic Policy Inference via Synthetic Observations