Leveraging Neural Koopman Operators to Learn Continuous Representations of Dynamical Systems from Scarce Data