Created by: sfraczek

This PR adds

• MKL-DNN support of Swish activation
• fuse of swish to conv2d.

This improves 1 core ShuffleNet performance by roughly 76% on Intel(R) Xeon(R) Gold 6248 CPU @ 2.50GHz

Test command:

./paddle/fluid/inference/tests/api/test_analyzer_int8_image_classification --batch_size=50 --infer_model /mnt/drive/PaddlePaddle/weights/generated_ShuffleNetV2_swish_layerwise --infer_data=/mnt/drive/data/ILSVRC2012/val.bin --iterations=200 --enable_int8=0

compared PR against this develop commit 99db0cf7

I generated the model by taking weights from https://github.com/PaddlePaddle/models/blob/develop/PaddleCV/image_classification/README_en.md#shufflenet-series the last one and adding save_inference_model(...) in eval.py from PaddlePaddle/models repo.