Reasoning with Sampling: Your Base Model Is Smarter Than You Think