Revisit CP Tensor Decomposition: Statistical Optimality and Fast Convergence