Upgrading Paddle SWIG API to C-API

[reyoung]

Here, we purpose several steps in developing Paddle API, make developing progress smooth and acceptable.

1. Use SWIG-API to expose Paddle and decide Python API.
   from 2016-12-20 to 2016-12-27

   1. give an example how to train a model in Python currently. @reyoung [2016-12-20]
      1. MNIST demo and Quick Start Demo.
      2. local mode / cluster mode.
   2. make trainer_config & data_provider in a same file. [2016-12-21, 22, 23]
      1. make trainer_config in the same python file. @jacquesqiao
      2. make data can easily fit to paddle gradient machine
   3. extract user level API from demo [2016-12-24,25]
      1. Add some syntactic sugar in python, make users not to write some standard code.
   4. change other demos by using API trainer. [2016-12-26,27]

2. Make Python-API stable, change it to C-API.

   1. We hope the Python-API is in alpha stage in 2016-12-27. So we can discuss API design base on this version. We will decide what Python API should be in early January.
   2. When the python API about Paddle is stable, we will remove SWIG and change it to C-API, let multiple language bindings could be possible.
      1. Clean Global variables/Remove log fatal.
      1. Expose C-API and make Python-API stable

3. Releasing C-API and SDK for C/Cpp model inference.