Improving Task-Specific Generalization in Few-Shot Learning via Adaptive Vicinal Risk Minimization