Uncovering the heterogeneous effects of depression on suicide risk conditioned by linguistic features: a double machine learning approach