Towards Accurate Post-training Network Quantization via Bit-Split and Stitching