Created by: baojun-nervana

1. close a TODO to remove the const_cast. This is to avoid altering the progdesc. Instead of substituting ngraph_engine op in progdesc, it is changed back to scan the op list after creation and replace the op sublist with ngraph_engine (the original way).

2. Refactor the engine code to reduce the overhead. The idea is to minimize the repetitives iteration to iteration. If the graph has no change, the NgraphEngine::Run will grab the data and run the computation.

   • an "engine_key" was added to verify if there is a graph change.
   • the repetitive info for NgraphEngine::Run will be cached