Steering LLMs Towards Unbiased Responses: A Causality-Guided Debiasing Framework