Fix #3386

* No POD instance is forbidden in global scope.

See
https://google.github.io/styleguide/cppguide.html#Static_and_Global_Variables

* merge InferShapeContext and OperatorContext

* OperatorBase& instead of OperatorBase*

* init grad op checker

* can run

* add GradeChecker class

* use get_numeric_gradient

* refine code

* add softmax and cross entropy auto grad test

* use close to judge op_grad and numeric_grad

* add cpu and gpu compare

* add comments

* add support_gpu

* fix allclose

* fix name error and symplify code

* optimize gradient checker

* add test_cross_entropy_op

* update gradient_checker.py

* optimize code

* use random.uniform instead of random.random

* fix type bug

* optimize check_grad

* put SupportGPU into OperatorBase

* typo

* fix some enforce

* remove compatible_type to avoid compile error

* remove shared_ptr

* fix tensor error msg

* finish forward debug

* init softmax grad op

* add compute code

* export Backward to python

* update test ,export op.type to python

* update python test, fix compute bug

* update unit test

* use eigen

* optimize eigen code

* add gpu test

* register softmax_grad GPU kernel and fix test bug

* typo

* follow comments

* Make interface of Operator to `const Scope&`

* use operator context

* optimize code

* update net infershape

* update InferShape

* disable override InferShape(scope) in OperatorBase

* change InferShapeImpl to InferShape

* add template to OperatorContext Input/Output

* merge Input InputVar, Output OutputVar

* change Inputs to MultiInput

* fix conflict

* fix MultiInput bugs and add unit test

* rename KernelContext to ExecutionContext

* clean code

* change InferShape to protected

* fix template bug

* refine code

* use InputVar instead of Input<Variable>

* typo

* optimize code

* 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.

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.