Outlier-Safe Pre-Training for Robust 4-Bit Quantization of Large Language Models