Spontaneous Reward Hacking in Iterative Self-Refinement