One embedding to predict them all: Visible and thermal universal face representations for soft biometric estimation via vision transformers