Thinking Slow, Fast: Scaling Inference Compute with Distilled Reasoners