GRAIT: Gradient-Driven Refusal-Aware Instruction Tuning for Effective Hallucination Mitigation