A novel risk-averse training framework that leverages score-based generative models for data augmentation tailored to Conditional Value-at-Risk minimization