Created by: yihuaxu

When run feature model, to use convolution operator with group attribute will get about 10% gain.

Platform: Intel(R) Xeon(R) CPU E5-2650 v4 @ 2.20GHz Command: ./paddle/fluid/inference/tests/api/test_analyzer_resnet50 --infer_model=pretrained_models/feature --gtest_filter=Analyzer_resnet50.profile_feature_mkldnn --paddle_num_threads=4 --repeat=1000 --batch_size=1

The below is the comparison with the original version.

Notice: As MKLDNN's convolution implementation, it need update the mkldnn version for AVX512 CPU. It should include the commit, the id is "c502e5f45410d9b2bb2155cb99364eaf2360c6db", in mkldnn.