RFG: Test-Time Scaling for Diffusion Large Language Model Reasoning with Reward-Free Guidance