Created by: shippingwang

Refactoring image classification. 分类代码重构

Background Following better coding rules mentioned in models new rules and to solve several issues and bugs

Include:

• Code(Script):

New train.py: Fewer code blocks(600l->200l) and better reading experience. Reader.py: Make batch data by paddle.batch now.

More details:

• Put all decay solvers and learning rate strategy in class: Optimizer
• Add some check functions in advance
• Integrate metircs in Loss, Include GoogLeNetLoss, MixupLoss, DefaultLoss, etc.
• Move print code to robust print function: print_info
• pyreader -> PyReader(iterable=Flase)
• PE -> compiled program
• New save/load methods
• Refine CE code, disable now
• Remove fp_16 and dist_train part, It will rewrite in PaddlePaddle/Fleet
• ngraph part, haven't test
• Add configs

主要改动：

• 重写代码，使train.py更加可读，包括pyreader->Pyreader, PE->compiled with形式，所有衡量指标放在Loss类中，所有学习率方式放在Optimizer类中 等。
• reader使用cv2作为默认处理

• Documents:

• move Readme.md to Readme_en.md and set Readme_cn.md(Readme.md) as default.

• update arguments description.

How to run this new code? Train: bash run.sh train AlexNet

fix #3262 (closed)