We put forward Op's inputs, outputs and output gradients into Grad
Op's inputs, and put forward Op's input gradients into Grad Op's output.
So Grad Op's `in_out_idx`, `input_format` and 'output format' need to be
rebuilt during Op creating.

* `in_out_idxs_` shares between all operator instance in same type
  of operator.

Net::DebugString() print what it is and invokes inner operators
DebugString() with a prefix '\t'.

* There is a merge conflict when merge PR #2914
* Develop and PR #2914 both add `DDim::size` method, but did not
  triger git merge conflict while merge.

* Follow review comments to seperate Tensor Numpy interactive methods in
  tensor.h.
* Simplify logic for `CastToPyBufferImpl`, make it as one struct and in
  details namespace.
* Remove `Scope` expose in Python, since it currently is useless.
* Remove some debug functions.

Basically following
http://pybind11.readthedocs.io/en/stable/advanced/pycpp/numpy.html

* Use buffer protocol to return a view of Tensor. It can be cast to
  numpy array in Python.
* Set a numpy array to a tensor.

* Also fix unit test

* add inputs

* add ut for multiple inputs

* fix AddToLayer

* op_desc -> op_proto

* CreateArgumentOffsetMap -> CreateInOutOffsetMap

* move CreateInOutOffsetMap from OperatorBase to op registry

* arg_idxs_ -> in_out_idxs_

* Implement InferShape and register them, give a stub Kernel method
  by LOG(INFO)

* add NDEBUG switch to PADDLE_ENFORCE

* using etcd as fault tolerant training

* update

* workable version, ft not tested

* small fix

* update

* remove TODO

* OperatorBase should not store OpDesc because not All op contains an
  OpDesc and not all ops create from OpDesc.
  * Networks do not contain OpDesc and are not created by OpDesc
* Do not register Network to OpRegistry.
  * The network is directly created by the user in Python. Not from
    registry.
* Correctly handle the `inputs` and `outputs` of a Network.
  * Add CompleteAddOp() methods
* Remove `AddOp(OpDesc&)` in net-op. All op are added by OperatorPtr.
* Rewrite unit test for truly tested what networks do.
* optimise operator_test

1. Refactor DDim's product() to make it more efficiently.

2. Add slice_ddim().

* add ADD_OP_CPU to enable add op with only cpu kernel

* It is used to create an operator library. It handles to split CPU and
  GPU sources and links operator common libraries.
* It also give a reasonable warning and error when operator developer
  not correctly implement an operator.
  * Warning for lack of GPU kernel.
* Same interface as `cc_library` to make code style consistent.

All OpCreation method are generated by
`create_op_creation_methods::__bootstrap__` method, and stores in
`op_creations` object and its methods.

There are three parts to implement this feature.

1. Get all registered `OpProto` from C++ side. It is implemented in
`get_all_op_protos` method.
1. Create a function to convert `kwargs` to `OpDesc` base on each op's
`OpProto`. The `OpDescCreationMethod` class.
1. Convert `OpProto` to `docstring` by `get_docstring_from_op_proto`
method.

All three methods are unit tested. The `__bootstrap__` just combines
them together and create a method in runtime.

For details, please reference the doc string in
`create_op_creation_methods.py` and the unit test
`test_op_creation_methods.py`.