Blas optimized elementwise_add forward and backward passes

[tpatejko]

This PR implements optimization of elementwse_add forward and backward passes.
It includes for forward pass:

• MKL VML-based optimization with v?Add then MKL/MKLDNN are used
• Blas-based optimization with VCopy and SAXPY operations when MKL is disabled

For backward pass:

• Blas level 1 VCopy is used for copying dx and dy vectors.

When integral or float16 types, or GPU device are used, the implementation falls back to the default (generic) elementwise_add operation.

[tpatejko]

This PR implements the following issue #10786.

[luotao1]

LGTM！Thanks for this speedup, I test it on OCR CRNN_CTC model, the total elapsed time (repeat 100 times of model) of elementwise_add op is from 467ms to 428ms.

[tpatejko]

@luotao1 Thanks for this information. Does the model converge?

[luotao1]

@tpatejko I only test the inference speed, but in our unit-tests, https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/fluid/tests/book will test the model converge.