Created by: reyoung

I just add a dependency engine to parse the dependencies of operators. There are still a lot of jobs need to be done.

1. ~Complete Broadcast parameters.~
2. ~Use thread pool to invoke operator parallelly.~
3. ~Complete NCCL AllReduce OpHandle in this implementation.~
4. Add fetch methods.

I just use VarHandle and OpHandle to parse Program as a SSA form graph. A variable is assigned by only one OpHandle. When all inputs of OpHandle is ready, the OpHandle can be run.

The speed of ResNeXt152 is