Consider making the Scan inner-graph an input to the Op

[brandonwillard]

Scan has a member/field that holds the inner-graph (i.e a graph representing the computation in the Scan's body). This design choice effectively hides the inner-graph from a standard graph traversal, and—as a result—forces the need for special Scan-specific logic changes in generic routines in order to work with these Ops.

We should consider changing this.