Contextual modulation of value signals in reward and punishment learning