fix bcast in reduce strategy

sync resnet model for fluid_benchmark

[MKLDNN] Fuse Conv+BatchNorm + ReLU

[MKLDNN] Reusing once reordered convolution weights in test mode

* add multiprocess_reader

* add multiprocess_reader to reader decorator

* support piped multi process reader

* revert v2 decorator

* add comment to multiprocess_reader

* optimize code

* use ujson to speed up json serialize/deserialize

* add assert to multiprocess_reader

* update comment of multiprocess_reader

* optimize ujson import, handle error case

* optimize import ujson

* remove ujson from requirements.txt

* add import sys to decorator.py

polish release doc

* add ignore index

* update api.spec

* enhance softmax_with_cross_entropy

refine Analysis test 

Fix cudnn holder bug

program and tensor versioning support

fix fluid benchmark script

Enable conv and batch norm by default

reset received vars on pserver

* Implement norm_op by CUDA instead of Eigen.
* Remove the commented code.