parallel-do should not merge the gradient of parameter that stop_gradient=True

Merge branch 'feature/parallel_nccl_not_support_batchnorm' of https://github.com/reyoung/Paddle into fix-parallel-do-with-non-parameter

Fix the profiler's bug in multi-gpu mode

Fix rename_var de-ref

refine FQA doc

* Register more data type for reshape operator.

* Enable the SSD loss to support normalization by the total number of output locations.

* Fix the doc format.

* Get rid of the dependency of Go compiler when WITH_GOLANG is OFF.

* Minor modification to the documentation.

* Remove the downloading and installing of Go compiler when build developing Docker image for android.

* Minor modifcation to the documentation's format.

* Change the default value of WITH_GOLANG to OFF in paddle/scripts/docker/build.sh.

Add cluster-train document.

Fix conv_op bug

remove `import paddle.fluid` in conf.py.in

combine batch_size_like.cc into batch_size_like.h

* merge fluid dist tests

* update cmake

add warrper for detection map operator

Update conll05.py

* Fix dist demo error

* revert trainer_id

Extend current profiler for timeline and more features.

Tensorflow benchmark

Refine cmake for cudnn op