Add memcpy_stream and use multi-thread to issue parameter copies

[panyx0718]

On TitanX with 4 devices without nccl, the se-resnext step
time reduces from ~1.18 to ~1.10

[chengduoZH]

It sounds great! but it may have a potential problem.
For example, one op_kernel has two operations, copying(GPU->GPU) and computation, if we put them into two streams(one stream is used to copy and the other is used to compute), before computation, the copy-stream must have a synchronize, or computation result maybe not correct.