Kernel Limit of Recurrent Neural Networks Trained on Ergodic Data Sequences