Learning salient representation of crashes and near-crashes using supervised contrastive variational autoencoder