hir typeck decision making now checks whether nested `eq`/`sub` obligations hold

[lcnr]

e.g. for coercions (rust-lang/rust#112022). This is necessary because we now emit nested obligations due to deferred projection equality and because we don't eagerly normalize anymore.

[compiler-errors]

i dont think we currently check nested obligations for InferCtxt::can_eq but we probably should do that there. the places where that fn is used are mostly diagnostic, so we'd need to check the non-diagnostic usages and consider the change in behavior.