ParetoQ: Scaling Laws in Extremely Low-bit LLM Quantization