Distribution learning via neural differential equations: a nonparametric statistical perspective