Language models can learn to tell when their activations are being read