Gravar e-mail: Neuron-by-Neuron Quantization for Efficient Low-Bit QNN Training