[Custom Descriptors] Optimize descriptors in Unsubtyping

[tlively]

With the upcoming stricter validation rules for descriptor types, we can
no longer use the placeholder describee trick in GlobalTypeOptimization,
so its ability to optimize out unneeded descriptors will be
significantly restricted. Because of the increasing coupling between the
validation for subtyping and descriptor relationships, it makes most
sense to optimize both at the same time in Unsubtyping.

Update Unsubtyping to start the analysis assuming no descriptors are
necessary, then add them in as it discovers that they are indeed
necessary either to preserve behavior or validitiy. Discovering new
required descriptors can imply that new subtypings are necessary and
vice versa.

[tlively]

All comments addressed except the one about the side effects of null descriptors in allocations.

[tlively]

Ok, now the traps are properly preserved. Will fuzz overnight.

[kripken]

Perhaps add a test for a non-nullable descriptor in a global?

[tlively]

This is tested toward the bottom with the comment ;; When we optimize out descriptors, we may need to update allocations..

[kripken]

Is there intentionally no arrow from top to top.desc?

[tlively]

Yes, the arrows in this diagram are only the initial edges required by the code.

[kripken]

Maybe that can be clarified? I was reading all these diagrams assuming they describe the situation in the wasm, and the wasm itself does contain a descriptor relationship there. I think that is what you've done elsewhere?

[kripken]

Maybe a good way to do this is to graph the initial wasm, and then put this second graph after it?

[tlively]

The "described by" arrows in the other diagrams are inherited from before this change, when we never removed descriptors, so they were irrelevant to the test cases. I'll update them to include only the initially required descriptor edges, just like they include only the initially required subtype edges.

[kripken]

lgtm otherwise