Label Smoothing Improves Gradient Ascent in LLM Unlearning