Gradient-free prox-methods with inexact oracle for stochastic convex optimization problems on a simplex