rustdoc: fix & clean up handling of cross-crate higher-ranked parameters
Preparatory work for the refactoring planned in #113015 (for correctness & maintainability).
---
1. Render the higher-ranked parameters of cross-crate function pointer types **(*)**.
2. Replace occurrences of `collect_referenced_late_bound_regions()` (CRLBR) with `bound_vars()`.
The former is quite problematic and the use of the latter allows us to yank a lot of hacky code **(†)**
as you can tell from the diff! :)
3. Add support for cross-crate higher-ranked types (`#![feature(non_lifetime_binders)]`).
We were previously ICE'ing on them (see `inline_cross/non_lifetime_binders.rs`).
---
**(*)**: Extracted from test `inline_cross/fn-type.rs`:
```diff
- fn(_: &'z fn(_: &'b str), _: &'a ()) -> &'a ()
+ for<'z, 'a, '_unused> fn(_: &'z for<'b> fn(_: &'b str), _: &'a ()) -> &'a ()
```
**(†)**: It returns an `FxHashSet` which isn't *predictable* or *stable* wrt. source code (`.rmeta`) changes. To elaborate, the ordering of late-bound regions doesn't necessarily reflect the ordering found in the source code. It does seem to be stable across compilations but modifying the source code of the to-be-documented crates (like adding or renaming items) may result in a different order:
<details><summary>Example</summary>
Let's assume that we're documenting the cross-crate re-export of `produce` from the code below. On `master`, rustdoc would render the list of binders as `for<'x, 'y, 'z>`. However, once you add back the functions `a`–`l`, it would be rendered as `for<'z, 'y, 'x>` (reverse order)! Results may vary. `bound_vars()` fixes this as it returns them in source order.
```rs
// pub fn a() {}
// pub fn b() {}
// pub fn c() {}
// pub fn d() {}
// pub fn e() {}
// pub fn f() {}
// pub fn g() {}
// pub fn h() {}
// pub fn i() {}
// pub fn j() {}
// pub fn k() {}
// pub fn l() {}
pub fn produce() -> impl for<'x, 'y, 'z> Trait<'z, 'y, 'x> {}
pub trait Trait<'a, 'b, 'c> {}
impl Trait<'_, '_, '_> for () {}
```
</details>
Further, as the name suggests, CRLBR only collects *referenced* regions and thus we drop unused binders. `bound_vars()` contains unused binders on the other hand. Let's stay closer to the source where possible and keep unused binders.
Lastly, using `bound_vars()` allows us to get rid of
* the deduplication and alphabetical sorting hack in `simplify.rs`
* the weird field `bound_params` on `EqPredicate`
both of which were introduced by me in #102707 back when I didn't know better.
To illustrate, let's look at the cross-crate bound `T: for<'a, 'b> Trait<A<'a> = (), B<'b> = ()>`.
* With CRLBR + `EqPredicate.bound_params`, *before* bounds simplification we would have the bounds `T: Trait`, `for<'a> <T as Trait>::A<'a> == ()` and `for<'b> <T as Trait>::B<'b> == ()` which required us to merge `for<>`, `for<'a>` and `for<'b>` into `for<'a, 'b>` in a deterministic manner and without introducing duplicate binders.
* With `bound_vars()`, we now have the bounds `for<'a, b> T: Trait`, `<T as Trait>::A<'a> == ()` and `<T as Trait>::B<'b> == ()` before bound simplification similar to rustc itself. This obviously no longer requires any funny merging of `for<>`s. On top of that `for<'a, 'b>` is guaranteed to be in source order.
3e293634e2