Stable elemwise mul

They were using float64 for FP32 kernel before.

Support double precision

* Adding support for the sigmoid_cross_entropy_with_logits operator

* Fixing a typo in the cuda file

* Adding Python documentation for sigmoid_cross_entropy_with_logits_op

* Correcting typos in documentation

* Adding unit tests for sigmoid_cross_entropy_with_logits_op

* Addressing code review feedback

Add reduce op

fix compile error

Use scalar implementation instead of neon implementation to avoid out of range memory access in the tail conv3x3

should reset env every time

move Protobuf desc to framework

Remove OperatorBase::InferShape

User StridedMemCpy in Concat/Split Op

Use scalar implementation instead of neon implementation to avoid out of range memory access in the tail conv3x3.

InferShape in Operator should be performed in OperatorBase::Run.

* cond_op, recurrent_op and mnist might be changed in following PR

Fixes #4445

Make PyBind support C++ exception