REAL: Reading Out Transformer Activations for Precise Localization in Language Model Steering