float16 related work by Kexin

[wangkuiyi]

1. float16 type definition and design doc: https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/float16.h
   https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/data_type/float16.md
2. Make some operators work with float16: https://github.com/PaddlePaddle/Paddle/issues?utf8=%E2%9C%93&q=is%3Aissue+float16+support+
3. Make the inference transpiler support float16: https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/inference_transpiler.py
4. Correctness verification reused https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_image_classification.py
5. Performance benchmarking: https://github.com/kexinzhao/Paddle/blob/float16_demo/python/paddle/fluid/tests/demo/float16/run_float16_demo.sh
6. Benchmarking report: https://github.com/kexinzhao/Paddle_benchmark/blob/master/float16_benchmark.md

[wangkuiyi]

The above work scatters around the Paddle repo. I'd suggest that we create a new folder:
/contrib/float16, and we move

3. inference transpiler (only the float16 part)
4. correctness verification and saving the transpiled VGG16 inference program
5. performance benchmarking, and
6. the report

into this folder.