Using double precision to stablize lstm gradient check

* init remove cond_op infershape

* optimize code

* add PrepareDataForSubnet and MergeDataFromSubnet

Few typos in refactorization file

Unify Activation functions and simplify register code

Fix row buffer mem free

Fix infer when input is empty in v2/inference.py

refine "Testing Backward Operators" in new_op_cn/en.md

fix typo

remove rowwise_add_op

fix random seeds in gradient checkers.

Add sequence softmax operator.

Pass DeviceContext to functors not ExecutionContext.

Implementing the SoftSign activation operator

Add grad_op_desc_builder

Stable elemwise mul

They were using float64 for FP32 kernel before.

Support double precision