"add nccl cmake enforce"

[dzhwinter]

To support Multi-GPU, we use NCCL library to do parameter/gradients integration and distribution. This PR adds an enforce macro helper module, cmake scripts to support NCCL. Cause NCCL library needs to compatible with different CUDA version, we use the DSO loading technical, same as other module cublas, cudnn does.

[QiJune]

Please add nccl as a cmake external project, just like pybind/eigen3/... And make ci pass first.

[QiJune]

We have to ways to introduce NCCL to paddle:

1. Make sure NCCL has been installed in CI machine first. Both .h and .so are installed in the correct directory. We can dynamic load .so of NCCL.
2. We download NCCL source code from github, and just take NCCL as a cmake external project. Since we fix the version of NCCL, we can build NCCL and use .a of NCCL in our project.

[dzhwinter]

NCCL2 has a good back compatible ability, which can used base CUDA 7.0 - CUDA 9.x.
Since the nvidia-docker image can install NCCL2 manually, so I will remove the github clone related cmake scripts.
nvidia-docker

NCCL is not bundled in any CUDA image right now, so you will have to install it manually for now. On 8.0 and 9.0.
Yes, I plan to add NCCL 2.0 to all images soon.
In the mean time, we have a new tag called "base" for CUDA 9.0. It adds our package repositories to the official ubuntu 16.04 image. Afterwards you can manually install any package:
FROM nvidia/cuda:9.0-base
RUN apt-get update && apt-get install -y libnccl-dev

[dzhwinter]

This will be closed since #5001 cherry-picked this code.