Provably Efficient Gauss-Newton Temporal Difference Learning Method with Function Approximation