Trainability, Expressivity and Interpretability in Gated Neural ODEs