[draft] a simple op design doc

[JiayiFeng]

No description provided.

[wangkuiyi]

Consider add bool is_tensor = 3;.

[wangkuiyi]

Copy https://gist.github.com/reyoung/dd1441659ef4806e0a1368f6fb8c6211#file-create_python_ops_functions-py into here.

[wangkuiyi]

paddle.layer?

[wangkuiyi]

This section "Op and kernel" need more discussion. We can remove this section from this PR, and merge it. So we could add this section in future PRs.

[wangkuiyi]

LGTM

[wangkuiyi]

Given

virtual bool Operator::ValidateAttrs() const {}

we don't need max and min here.

[wangkuiyi]

Work in parallel:

1. Given OperatorProto generate Python op creator functions.
2. Python op creator functions calls paddle::framework::CreateOp.
3. Python op creator functions call C++ functions via C wrappers.
4. paddle::framework::CreateOp depends on operator registration.

   class Operator {
    public:
     ~Operator() {}
     virtual std::string ValidateAttrs() { return validator_(); }
     virtual void Run() = 0;
   
    private:
     std::function<std::string()> validator_;
   };