After #20910, data parallel implementation still needs the following improvements:

• Fix C++ data parallel for BN
• Make sure C++ data parallel work for double backward
• Move reduce_add and reduce_add_coalesced in torch/cuda/comm.py to C++.
• Move Broadcast and ReduceAddCoalesced from torch/nn/parallel/_functions.py to C++.
• Make C++ data parallel use ReduceAddCoalesced.
• Consolidate C++ and Python module replicate.