Experiment on vgg flower, 2 trainers, 1ps.
    more trainer could have more speedup.

    After:
    Pass = 0, Iters = 327, Speed = (7.52) img/s
    Before:
    Pass = 0, Iters = 385, Speed = (6.77) img/s

Async GRPC sendrecv

* send_recv variable

* delete unused logs

* fix ci failed

* update

* resize tensor before tensor copy

* add selectedrows unit test

* check rows

* add back print_operators_doc

* fix style check

* fix style check

* Add Python wrap of conv2d_transpose and its unittest

* Follow comments

* Fix format

* WIP send recv op

* WIP send recv

* put grpc impl in details

* put grpc impl in details

* update wip

* update proto

* update proto

* update proto

* clean cmake

* wip on op implementations

* wip on op implementations

* compile ok adding ut

* wip unitest

* add extern cares for linking

* wip add ut

* working version send recv

* revert optimizer.py

* update test cmake

* add libtool to dockerfile

* update cmake dependency

* update cmake depends

* update cmake grpc depends

* fix cmake dependency

* fix compile error

* fix compile

* follow comments

* update

* update copyfrom

* Use stable_sort in lod_rank_table

It is easy to debug and test when use `stable_sort`and the time
complexity is not changed.

* Add LoDTensorArray

* Feed/Fetch op just plain operator, not a OpWithKernel
* Do not register OpInfoMaker since Feed/Fetch will never be
  configured by users
* Feed/Fetch op has empty gradient
* Feed/Fetch op do not hard code `feed_variable`, `fetch_variable` as
  its input and output, make it as a plain Operator input/output

* add NDEBUG switch to PADDLE_ENFORCE