AIs that tackle maths’ hardest problems may herald human-level reasoning. Can they pave the way to artificial general intelligence, asks Alex Wilkins