We need to refactor/rewrite the graph caching logic. We will be implementing a 2 stage graph cache system. One that is based on fusion when contiguity, type, or device changes. The second is related to heuristics.

This first graph cache system should be pretty straight forward to reimplement but I'd like to think this back through based on the overall system, and make sure there's a clear reuse, and separation with the heuristic caching logic.