* 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_

* add NDEBUG switch to PADDLE_ENFORCE

* 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

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`.

1. Add template T which indicates data type to `CopyFrom()`, `Slice()`
and `ShareData()` functions. This makes `CopyData()` code much clearer.

2. Add `set_dim()`.

3. `product(DDim)` transforms `DDim` to `vector<int>` first and then calculate
its product. That might be quite slow. For `product(dims_)` is frequently
used in Tensor, we add a mumber variable `numel_` as a cache of the
product result.
TODO: refactor `product()` to make it more efficient.

4. Unable Tensor::operator=

5. Remove the limit of POD type, because `float16` and `int8` are not POD type.

* use OperatorPtr = std::shared_ptr<OperatorBase>;
* use ScopePtr = std::share_ptr<Scope>;

* Let OpProto support multiple and temporary

* Each input/output of Paddle's Op could be a list. Add multiple mark to
  OpProto. Also add a `input_format`/`output_format` attribute if that
  Op has multiple input or output. The format of that attribute please
  reference the comments in `op_proto.proto`
* Add temporary mark, because some output of an Op is not used by user
  but used by other op for faster computation. Explicitly mark which
  output is temporary could let future memory/computation optimization.
* Add generated field to AttrProto.

* Add `AddInputs`/`AddOutputs` function

* It is more readable to invoke `AddInputs` not
  `AddInput(multiple=true)`.

* Convert `op` --> `operators`
* Remove AddType in OpProtoMaker, because type is part of registry.
* Rename CPU_OR_GPU --> DEVICE_TYPE in registry macro.

* Refine register methods, make Op can get rid of whole-archieve
* `USE_OP` before a op is used.
* Add unittest for add_op.

Add unit test for OpKernel

1. Add `Tensor::CopyFrom`. Current version can only support CPU memory
copy. The support of GPU will be provided later by `paddle::memory`.
The current implementation of `Tensor::CopyFrom` is a little inefficient:
Every time `CopyFrom` is called, tensor will re-allocate its memory. However, if
we try to check and reuse `placeholder_`, we have to provide a template
parameter for `CopyFrom` to indicate the data type. It seems strange for
a simple copy function.

2. Add `Tensor::mutable_data(Place place)`, which directly use member
variable `dims_` as its dim parameter. This interface is required by
`Op::InferShape`.

* User can register OpKernel to its Ops. The OpKernelMap saved in
  OperatorWithKernel. Each Op which inherits OperatorWithKernel will
  use `OpKernel::Compute` instead of Run.

Paddle's data block is row-major order, while Dim::contiguous and
Dim::contiguous_strides are based on column-order. So remove them to
prevent misuse.

Add OperatorBase.

issue: https://github.com/PaddlePaddle/Paddle/issues/2790

Paddle design the Operator with Kernel. OperatorBase has no type and device information when create, One operator can have multiple kernels, Operator will choose a kernel to run according to context. The kernel should be bind to Operator before or during Operator running.

1. Add member variable 'DDim dims_' and a getter function 'dims()'.
'dims' is supposed to hold tensor's shape during Op::InferShape.
2. Remove 'mutable_data' which use default Place. User must specify a
explicit Place when call 'mutable_data'.
3. A PlaceHolder may be shared by more than one tensor, and some of them may be the others' slices. So we add a new member variable 'offset_' for Tensor, which is used to show the byte offset between PlaceHolder::ptr_ and where tensor's data really begins.
4. Add functions 'ShareDataFrom' and 'Slice' for Tensor.

TODO: Tensor needs a 'CopyFrom' function.

* Move static variable defined in .cc

We cannot define static variable in .h, because it will be multi-defined
errors.

Also fix some cpp syntax, like:

* Prefer to use algorithm not manually for-loop, to make code more
  readable.
* Remove unused `()`.
* Enforce take a bool. It is no need `xxx==true`.
* Use range-based for-loop iterator from op_desc.attrs

* Fix a protential static variable init order error