Created by: liangan1

Conv + Sigmoid fusion is a key optimization for Squeeze-and-Excitation Networks (https://arxiv.org/pdf/1709.01507.pdf). The representative model is SE-ResNext and its variant. With this fusion, we can reduce the memory footprint for Conv + Sigmoid and therefore improve the performance. As MKL-DNN supports the primitive post op in FP32, we submit the PR to support FP32 first and will add INT8 once MKL-DNN provides the support.