On convergence of a $q$-random coordinate constrained algorithm for non-convex problems