Distribution learning via neural differential equations: minimal energy regularization and approximation theory