Causal-Informed Contrastive Learning: Towards Bias-Resilient Pre-training under Concept Drift