There are three benefits:
1. GetModIfaceFromDisk and GhcSessionDeps no longer depend on the transitive module summaries. This means fewer edges in the build graph = smaller build graph = faster builds
2. Avoid duplicate computations in setting up the GHC session with the dependencies of the module. Previously the total work done was O(NlogN) in the number of transitive dependencies, now it is O(N).
3. Increased sharing of HPT and FinderCache. Ideally we should also
   share the module graphs, but the datatype is abstract, doesn't have a
   monoid instance, and cannot be coerced to something that has. We will
   need to add the Monoid instance in GHC first.

On the Sigma repo:
- the startup metric goes down by ~34%.
- The edit metric also goes down by 15%.
- Max residency is down by 30% in the edit benchmark.

Surfacing the performance tradeoffs in the core build rules