regression when relating two opaque types in canonical queries

[aliemjay]

Each revision of this test is an ICE in 1.75.0-nightly (2023-10-17 09df610):

// revisions: sized clone

#[cfg(sized)] fn rpit() -> impl Sized {}
#[cfg(clone)] fn rpit() -> impl Clone {}

fn same_output<Out>(_: impl Fn() -> Out, _: impl Fn() -> Out) {}

pub fn foo() -> impl Sized {
    same_output(rpit, foo);
    same_output(foo, rpit);
    rpit()
}

Both revisions have a somewhat convoluted history best shown in a table form:

	sized	clone
stable 1.73	PASS 🆗	ICE 🧊
#114586 (1.74.0-beta)	PASS 🆗	PASS 🆗
#116530 (1.75.0-nightly)	ICE 🧊	ICE 🧊

While it may seem like #116530 is the culprit, I'd argue that the PR #114586 introduced an incomplete fix that I don't believe we should commit to in stable. Additionally, reverting #114586 while keeping #116530 would fix sized version as well as an other regression #114586 (comment), making it the right choice in my opinion.

[apiraino]

WG-prioritization assigning priority (Zulip discussion).

@rustbot label -I-prioritize +P-critical

[aliemjay]

#116879 was backported to beta; clone revision of the test above no longer passes in beta and sized revision is no longer an ICE in nightly.