Rollup merge of #140276 - compiler-errors:typeof-less-eagerly, r=lcnr

Do not compute type_of for impl item if impl where clauses are unsatisfied

Consider the following code:

```rust
trait Foo {
    fn call(self) -> impl Send;
}

trait Nested {}
impl<T> Foo for T
where
    T: Nested,
{
    fn call(self) -> impl Sized {
        NotSatisfied.call()
    }
}

struct NotSatisfied;
impl Foo for NotSatisfied {
    fn call(self) -> impl Sized {
        todo!()
    }
}
```

In `impl Foo for NotSatisfied`, we need to prove that the RPITIT is well formed. This requires proving the item bound `<NotSatisfied as Foo>::RPITIT: Send`. Normalizing `<NotSatisfied as Foo>::RPITIT: Send` assembles two impl candidates, via the `NotSatisfied` impl and the blanket `T` impl. We end up computing the `type_of` for the blanket impl even if `NotSatisfied: Nested` where clause does not hold.

This type_of query ends up needing to prove that its own `impl Sized` RPIT satisfies `Send`, which ends up needing to compute the hidden type of the RPIT, which is equal to the return type  of `NotSatisfied.call()`. That ends up in a query cycle, since we subsequently try normalizing that return type via the blanket impl again!

In the old solver, we don't end up computing the `type_of` an impl candidate if its where clauses don't hold, since this select call would fail before confirming the projection candidate:

https://github.com/rust-lang/rust/blob/d7ea436a02d5de4033fcf7fd4eb8ed965d0f574c/compiler/rustc_trait_selection/src/traits/project.rs#L882

This PR makes the new solver more consistent with the old solver by adding a call to `try_evaluate_added_goals` after regstering the impl predicates, which causes us to bail before computing the `type_of` for impls if the impl definitely doesn't apply.

r? lcnr

Fixes rust-lang/trait-system-refactor-initiative#185

Author: ChrisDenton