Crash training, if the number of samples is less than the count of devices.

* add cuda_device_functions.h

* move reduceSum to elementwise_op_function.h

Fix dist transpiler bug

fix python package have no version.py

Simplify DataStructure in SSAGraph

Add title for kernel_hint_design.md & kernel_selection.md

Update build and install/index_en.md

Avoid init_p2p all the times

Add multi-thread inference example which shares the inference_program and parameters

Rearange parallel_executor.md

Fix some dead links for fluid documents (cn version)

Add some comments for distribute_transpiler

Fix CPPlint issues in some operators

Fix CPPLint issues in spp_op, sum_op, topk_op, transpose_op, unpool_Op and warpctc_op

* Remove call_once.h

* "fix ci"

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into feature/simplify_data_structures

Add gitignores

add insert_op for block