White-box Testing of NLP models with Mask Neuron Coverage