Towards Tracing Trustworthiness Dynamics: Revisiting Pre-training Period of Large Language Models