separate the data parallelization from model parallelization

[bobye]

change backpropagate() to two versions (one is sequential in data, one is parallel in data)

[bobye]

Another workaround is to consider pass derivative as explicit outputs, and use aggregate to obtain overall gradient.

[bobye]

I like the second solution to handle data parallelization: It always keeps the status immutable.