On Computability, Learnability and Extractability of Finite State Machines from Recurrent Neural Networks