Created by: baojun-nervana

This PR includes a number of changes to enable BERT through ngraph:

1. set ngraph to support block 0 only, the subblock from condition_block op will run through paddle
2. Infershape was delayed before getting output dimension.
   • reason for this change is some of the variables are reused in some reshape ops and it will impact the inputs when the inputs trying to pull the shape.
   • mul_op and pool2d were updated accordingly per this change
3. fix some corner cases
   • handle shape {0}
   • partial support of lookup_table_grad
4. remove duplicated output
   • this is an optimization
5. Handle reshape2_grad infershape Paddle reshape2 add a leading dimension 0 and chop it off in reshape_grad infershape. Ngraph would interpret 0 as 0 element, so it needs special handling in ngraph engine