Created by: qingqing01

1. Now only support GPU, CPU will be add in nexted PR.
2. Use ncclAllReduce to sync E(x) and E(x^2) on multi-gpus in one machine. 3. Add ncclComm_t in DeviceContext and initialized by init.cc
3. The unit testing is to compare the forward outputs and backward outputs with batch_norm on one GPU.

build_strategy = fluid.BuildStrategy()
build_strategy.sync_batch_norm = True
binary = fluid.compiler.CompiledProgram(tp).with_data_parallel(
        loss_name=loss_mean.name,
        build_strategy=build_strategy)