Higher-ranked subtyping is too strict #33684

Branching off from #32330. The current subtyping rules for higher-ranked regions are overly conservative, with the result that a function declaration like this:

fn foo<'a>() -> &'a i32

is not equivalent to

fn foo() -> &'static i32

This exaggerates the impact of #32330 because functions like

fn foo<'a>(&self) -> &'a i32

must be converted to the form

fn foo<'a>(&'a self) -> &'a i32

in some cases, which feels a bit silly.

I proposed a fix in #32330 (described in some detail here), however I decided to hold off and land that fix separately, when I had more time to polish it (and consider it). In the meantime, this does mean that we wind up with more warnings than we otherwise would.

There is some mild backcompat risk in that the current coherence rules will allow you to distinguish between types on the basis of bound regions, but in some cases that is wrong. It is probably worth modifying coherence so that it doesn't distinguish bound regions at all -- or better yet, operates with fully erased regions.