Efficiently learning fermionic unitaries with few non-Gaussian gates