Efficient non-linear model reduction via a least-squares Petrov-Galerkin projection and compressive tensor approximations