uniquify root goals during HIR typeck

[lcnr]

We need to rely on region identity to deal with hangs such as rust-lang/trait-system-refactor-initiative#210 and to keep the current behavior of fn try_merge_responses.

This is a problem as borrowck starts by replacing each occurrence of a region with a unique inference variable. This frequently splits a single region during HIR typeck into multiple distinct regions. As we assume goals to always succeed during borrowck, relying on two occurances of a region being identical during HIR typeck causes ICE. See the now fixed examples in rust-lang/trait-system-refactor-initiative#27 and #139409.

We've previously tried to avoid this issue by always uniquifying regions when canonicalizing goals. This prevents caching subtrees during canonicalization which resulted in hangs for very large types. People rely on such types in practice, which caused us to revert our attempt to reinstate #[type_length_limit] in #127670. The complete list of changes here:

• new solver: implement canonicalization and region constraints #107981
• don't uniquify regions when canonicalizing #110180
• Restore region uniquification in the new solver 🎉  #114117
• add caching to most type folders, rm region uniquification #130821

After more consideration, all occurrences of such large types need to happen outside of typeck/borrowck. We know this as we already walk over all types in the MIR body when replacing their regions with nll vars.

This PR therefore enables us to rely on region identity inside of the trait solver by exclusively uniquifying root goals during HIR typeck. These are the only goals we assume to hold during borrowck. This is insufficient as type inference variables may "hide" regions we later uniquify. Because of this, we now stash proven goals which depend on inference variables in HIR typeck and reprove them after writeback. This closes rust-lang/trait-system-refactor-initiative#127.

This was originally part of #144258 but I've moved it into a separate PR. While I believe we need to rely on region identity to fix the performance issues in some way, I don't know whether #144258 is the best approach to actually do so. Regardless of how we deal with the hangs however, this change is necessary and desirable regardless.

r? @compiler-errors or @BoxyUwU