Avoid global trait caches while defining opaque types

[dominic-r]

Fixes #159932.

The old trait solver stores evaluated obligations in a global cache so they can be reused across inference contexts. While type checking a body that defines opaque types or coroutines, it avoided the global cache only when the current predicate mentioned one directly.

However, that check was insufficient. A predicate can depend on an opaque type through another obligation without mentioning the opaque type itself. A small source change could alter the order in which obligations were evaluated and therefore change the contents of the global cache. On the next incremental build, rustc could evaluate the same query key with different cache state and produce a different result. Incremental verification detected the changed fingerprint and raised the ICE.

So, this PR removes that exception. While a body defines opaque types or coroutines, the old trait solver now uses only the inference context's local caches. (These still avoid repeated work during the current type-checking run, but results are no longer shared between inference contexts where they may depend on an opaque type being defined.)

Do note that this may reduce global cache hits for async code and return-position impl Trait. It also removes part of the caching compromise introduced in #132625 for the compile-time regression tracked in #132064.

[rustbot]

Thanks for the pull request, and welcome! The Rust Project has assigned @mati865 (or someone else) to review your changes, you should hear from them (or someone else) within the next two weeks.

Please see the contribution instructions and our LLM policy for more information.

Why was this reviewer chosen?

The reviewer was selected based on:

• Owners of files modified in this PR: compiler
• compiler expanded to 75 candidates
• Random selection from 18 candidates

[mati865]

Not familiar with this area. @rustbot reroll