Assessing Robustness of Text Classification through Maximal Safe Radius Computation