PopuLoRA: Co-Evolving LLM Populations for Reasoning Self- Play