MGCA: require #[type_const] on free consts too

[khyperia]

When investigating another issue, I discovered that following ICEs (the const_of_item query doesn't support non-type_const-marked constants and does a span_delayed_bug):

#![feature(min_generic_const_args)]
#![allow(incomplete_features)]

const N: usize = 4;

fn main() {
    let x = [(); N];
}

My initial thought of "only require #[type_const] on places that stable doesn't currently accept" ran into the issue of this compiling on stable today:

trait Trait {
    const N: usize;
}

impl<const PARAM: usize> Trait for [(); PARAM] {
    const N: usize = PARAM;
}

fn main() {
    let x = [(); <[(); 4] as Trait>::N];
}

Figuring out which specific cases are not currently accepted by stable is quite hairy.

Upon discussion with @BoxyUwU, she suggested that all consts, including free consts, should require #[type_const] to be able to be referred to. This is what this PR does.

---

The change to tests/ui/const-generics/generic_const_exprs/non-local-const.rs is unfortunate, reverting the fix in #143106 no longer fails the test. Any suggestions to test it more appropriately would be most welcome!

edit: never mind, figured out how compiletests work :) - verified that the new test setup correctly ICEs when that PR's fix is reverted.

r? @BoxyUwU

[rustbot]

HIR ty lowering was modified

cc @fmease

[BoxyUwU]

Ah interesting, I'd have thought you'd have needed to update AssocConst too but lower_resolved_assoc_const_path already handles it. That's fun

[BoxyUwU]

Suggested change

	if find_attr!(self.tcx().get_all_attrs(def_id), AttributeKind::TypeConst(_)) {
	if self.tcx().is_type_const(def_id) {

should work instead and be a bit shorter/nicer :3

[BoxyUwU]

pointing to the definition site of the non-type const is rly nice :3 thanks