Merge branch 'support_scalar_kernels' of https://github.com/Xreki/Paddle into support_scalar_kernels

2. Centralize the use of sse and neon instrinsic.
3. Disable neon intrinsic when enable gpu.

Set protobuf version when PROTOBUF_ROOT set

remove duplicate code of trainer_config_helpers/config_parser.py

Fix baidu gcc toolchain compile

Fix add libn_any to dependencies

* It will cause build error when concurrent is large.

Remove not necessary functionalities in Parameter

Add with_testing guard for python/CMakeLists.txt

Add user can define PROTOBUF_ROOT.

Support variable-dimension input feature for 2D convolution operation.

Expose the "reverse" argument for recurrent_group in V2 API.

update design doc: Client Library of Parameter Server

implement task handling for master server's service

Allow setting both is_static and initial_mean, etc. at the same time

design doc for implementation parameters in CPP.

Add design doc for DeepSpeech2 on PaddlePaddle.

Some time we simply want to have a fixed initialization for a model.