Can Your Model Tell a Negation from an Implicature? Unravelling Challenges With Intent Encoders