Make some MIR ref types closer to the source they are lowered from

[ShoyuVanilla]

cc #149973 (comment)

r? RalfJung
Sorry for bothering you, but I think it might be better to get the required MIR changes approved first, and then move on to implementing the actual lints for recursive static refs 😅

[rustbot]

Some changes occurred to MIR optimizations

cc @rust-lang/wg-mir-opt

[ShoyuVanilla]

This wasn’t discussed in
#149973 (comment),
but I think it would be better to syntactically track promoted consts.

For example, the promoted const in the following code:

static FOO: () = {
    let _ = &&FOO;
};

is currently lowered to:

const FOO::promoted[0]: &&() = {
    let mut _0: &&();
    let mut _1: &();
    let mut _2: &();

    bb0: {
        _2 = const {alloc1: &()};
        _1 = &(*_2);
        _0 = &_1;
        return;
    }
}

I think it would make more sense for _2 to be immutable. This would also make it clearer when detecting &STATIC from:

_2 = const {alloc1: &()};
_1 = &(*_2);

It seems very unlikely that _2 would ever be truly mutable (for example, being reassigned with a different value), but anyway _2 being immutable gives some more assurance.