Enhancing Language Model Factuality via Activation-Based Confidence Calibration and Guided Decoding