optimize batch normalization and relu layer

[hjchen2]

Merged the BatchNorm and Scale layer.
Optimize most elapsed time of batch normalization by re-writing the forward and backward process using OpenMP rather than BLAS to avoid additional multiply operation.
The elapsed time of ResNet-50 with batch size set by 128 fells from 32523.77ms to 16982.93ms on Intel(R) Xeon(R) CPU E5-2450 0 @ 2.10GHz.

[jczaja]

Thanks very much we will have it reviewed shortly!

[tpatejko]

Thank you for your effort.

We reviewed the pull request and we are not going to merge it right now.

The code changes functionality of Intel Caffe:

• If I understand correctly, the pull request implements shift and scale in batch normalization. Original implementation separates these two operations;
• The pull request removes code guarded by variable use_global_stats. It might have a negative impact on the performance of trained network.