OLD: op_type -> grad_op_creator

NEW: grad_op_type -> grad_op_creator
     op_type -> grad_op_type

Replace `OperatorPtr` with `std::shared_ptr<OperatorBase>`

* ScopePtr means pointer of scope, but it can be shared or uniqued.
Change it to std::shared_ptr<Scope> to make code better to read.

* Expose Net to Python

* Expose PlainNet to Python, make python can add_op, complete_add_op
* Provide a low level api to manipulate Net
* Unittest for Net::DebugString

1. Change PODDeleter's template parameter 'PlaceType' to 'Place'.

2. Limit PODDeleter and Tensor::mutable_data()'s `T` to POD type.

ATTENTION: some interfaces changed:
1. void Tensor::set_dims(const DDim& dims) ==> void Tensor::Resize(const DDim& dims).
2. void Tensor::ShareDataFrom(const Tensor& src)  ==> void Tensor::ShareDataWith(const Tensor& src)
3. DDim Tensor::dims() const ==> const DDim& Tensor::dims() const

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.