Causal Graph Guided Steering of LLM Values via Prompts and Sparse Autoencoders