Towards Efficient Pre-training: Exploring FP4 Precision in Large Language Models