* Add design of switching kernel

* Follow comments

* multi cpu design

* update

* multi cpu executor to executor

* add graph converting

* use parallel operator to execute blocks with multi threads

* use auto-transpiler

* use auto-transpiler

* update

* update graph

* Polish executor design doc

* Adding few details

* Add fluid.md

* in response to comments from Tony and Abhinav

* add support new device doc

* add more info

* write versino.py

* add version py

* clean init py

* add istaged, major and etc... fields

* update

* update

* update

* Add survey of support of half in different CUDA versions

* small fix

* Updating the writeup of RNN doc

* Editing the documentation for seq_decoder, and fixing typos

* Fixing the captioning on 2 level RNN

* pushing after a pull

* Editing and re-writing parts of Data Reader design doc

* Chage `IndicateDataType` to `GetKernelType`. Make it easier to
  understand.
* Change `OpKernelKey` to `OpKernelType`
* Make operator developers can customize which kernel the operator will
  use in runtime.