Draft for train by parallel executor

improved the documentation for sequence_pool function

Fix signed/unsigned comparison warning

Support nccl2 dist train in trainer

benchmark/fluid script支持多卡训练

add pybind.h to inference lib for static link

Make tensor support uint8

Refine fetch op handle

Polish cmake

* fix build capi

* update by comment

* "fix ci"

* "fix network config"

* "rerun ci"

code clean