Globally Convergent Accelerated Algorithms for Multilinear Sparse Logistic Regression with $\ell_0$-constraints