HAVE: Head-Adaptive Gating and ValuE Calibration for Hallucination Mitigation in Large Language Models