Contracting Implicit Recurrent Neural Networks: Stable Models with Improved Trainability