On the Computability of Robust PAC Learning