DenseShift: Towards Accurate and Efficient Low-Bit Power-of-Two Quantization