LOST: Low-rank and Sparse Pre-training for Large Language Models