Investigating Bias Representations in LLMs via Activation Steering