Nash CoT: Multi-Path Inference with Preference Equilibrium