Fix Leaf Ops in Graph

Add CUDAPinnedPlace

All leaves must be variables. When all variables are ready, the
execution will be completed. If a operator has no output, the `Op::Run`
might not be started when the execution of graph has been complete.

Fix cpplint errors operators detail

Implementation of MKLDNN FC

Add upsample layer

prepare pserver executor

Fix compilation warnings in variable_response.cc

Fix compare warning in listen_and_serv_op.cc

Merge branch 'develop' of https://github.com/paddlepaddle/paddle into fix_cpplint_errors_operators_detail

* Update recurrent_group_en.md

Fix https://github.com/PaddlePaddle/Paddle/issues/9013

* Update recurrent_group_en.md

refine sync_with_cpp when remove ops or remove vars

Group nccl all reduce and improve performance (~14% for 4 device resnext)

Add python wrapper for pad_op

Fix table display errors for `.md` files