Fix shape inference bug

[reminisce]

This PR fixes a shape inference bug in function InferAttr.
Consider the following network:

weight --> abs_op --> sum_op 
       \                    \ 
        \                     add_op
         \                  /
data   --> fc_op --> sum_op

Given a data shape, if the shape forward inference starts from weight, there would be several node entries with unknown shapes for abs_op and sum_op in the first pass. The weight shape is only known after fc_op's shape inference is done. When the backward shape inference starts, it assumes that all forward node entries' shapes are known. However, this is not true in this case and therefore, leads to CHECK_EQ failure.

The problem is fixed by skipping CHECK_EQ for empty shapes in the backward shape inference.

@piiswrong @eric-haibin-lin

[tqchen]

@reminisce Can you also backport this to nnvm?

[reminisce]

@tqchen I will do that. Thanks.