On complexity and convergence of high-order coordinate descent algorithms for smooth nonconvex box-constrained minimization