Tracking issue to remove `TyKind::CoroutineWitness` from the type system

I'm interested in removing the TyKind::CoroutineWitness variant from the type system.

Coroutine witnesses are a synthetic generic arg given to coroutines to stall the computation of the "witness" types of the coroutine (which are the variables live across await points) by initially being populated with an infer var, then getting inferred to a tuple after analysis was run at the end of typeck. (We do similar things for closures, for example, for their upvars.)

Since #101692, the tuple was replaced with a TyKind::CoroutineWitness which defers that computation further to MIR. However, we still use an infer var to stall obligations that would rely on using these witness types (and only unify with with a CoroutineWitness at the end of typeck), since if we try to query for the witnesses during typeck in the same body that defines the coroutine, we would encounter query cycles.

#138845 introduced a more appropriate way to stall oblgiations that rely on coroutine witnesses which is currently used in the new solver. I'd like to use this in the old solver, and eventually rip out the TyKind::CoroutineWitness altogether, since it's redundant with TyKind::Coroutine which stores the same def-id and args as the former.

This is a tracking issue to make sure that I follow up with the work.

Rough timeline:

• Track coroutines that must be stalled in TypingMode: https://github.com/rust-lang/rust/pull/138845
• Stop using infer vars to stall coroutine witnesses in old solver: https://github.com/rust-lang/rust/pull/141762
• Fix the way we handle coroutines in drop live outlives machinery
• Remove TyKind::CoroutineWitness
• profit

Start by reviewing the issue's completed work in PRs #138845 and #141762, then investigate the unchecked coroutine handling in drop live outlives machinery. The goal is to fix that handling and eventually remove TyKind::CoroutineWitness, verifying that coroutine type-system behavior remains correct.