Explainable machine learning for mental health prediction from social media behavior: a nested cross-validation study with SHAP and LIME interpretability