Streamline the graph evaluation process to reduce overhead.

Important changes / gotchas:
* get_cached_region is no longer used. The cache bounds are
  always the full output bounds, but the request will not be
  expanded.
* The cache intersect is calculated after all requests to a node
  are combined, so partial cache hits from multiple nodes may
  skip the cache.
* Meta nodes (nodes who's pads do not reference them) will never
  be visited.
* Prepare is called only once per node, and must not modify the
  connectivity of the graph.