2. refine trt code
test=develop

* refine fused_all_reduce_op

* add unit test in test_parallel_executor_seresnext
test=develop

Fix topk shape check

fix squeeze shape check

Fix sequence pad 2

fix matmul shape check

* Rewrite gradient ProtoMaker for affine_channel_op to remove the Output as the input.
* Add act in Python API to make the act can be in-place by layer_help.py

remove graph visualizer tool, which can be replaced by python IrGraph draw API

Fix lod reset

Approval rd api

* temoprarily disable the code of use kCUDNN, test=develop

* add TODO comment, test=develop

fix formula in dropout

* add resnet nccl2 test=develop

* test dist train test=develop

* update test=develop

* increase timeout test=develop

* test on CI env test=develop

* Add enabling quantization

test=develop

* remove unused (here) function

fix pattern maching conv2d with(out) ResidualData

* Enhance affine_channel_op infer-shape check

* save infer model bug fix, return target vars' name list

not use PERSISTENT in batch_norm. test=develop

1. refine anakin engine
2. add data type for zero copy

align dev branch and PaddlePaddle:feature/anakin-engine brach
the cudnn workspace modify was not included for now, because we use a hard code way
in feature/anakin-engine branch. There should be a better way to implement it,
and subsequent submissions will be made.

test=develop

support change input size

* use anakin batch norm and scale implement fluid batch norm

cherry-pick feature/anakin-engine: add anakin softmax/transpose/batch_norm/flatten/reshape op (#16020)

* add anakin softmax/ flatten/reshape/transpose/batch_norm op converter