rustdoc doesn't render non-trivial built-in `Sized` impls (and calls `Sized` impls auto trait impls)

[fmease]

Since #78181, we leverage our auto trait impl synthesizer to render negative impls of Sized for types that don't implement Sized. Contrary to actual auto traits, we intentionally never render any positive Sized impls, only negative ones or none at all. Most ADTs are unconditionally Sized, so it makes sense to omit these to avoid unnecessary visual clutter and to reduce page size.

Quick aside: Sized is not an auto trait, so ideally synthetic Sized impls wouldn't appear under a section titled Auto Trait Implementations but instead under one titled Built-in Trait Implementations (modulo bikeshedding).

I take issue with the fact that we don't render "non-trivial" built-in Sized impls. Why bother? Well, Sized is fundamental (so making a type Sized is a breaking change) and under which conditions the last field of a struct is Sized is technically part of the public API and should thus be accessible in rustdoc-generated docs.

If rustdoc@main doesn't render a Sized impl for certain generic types, they are either unconditionally or conditionally sized but sadly no further information is provided. Consider these examples:

• struct Type<X: ?Sized>
  • is Type<T> Sized for all types T? holds if body is e.g., (Box<T>)
  • is Type<T> Sized iff T is Sized for all types T? holds if body is e.g., (T)
• struct Type<X: Deref>
  • is it unconditionally sized?
  • is it Sized iff <T as Deref>::Target is Sized for all types T?
• struct Type<X: ?Sized, Y: ?Sized>
  • which type param is responsible if any?
• [and so on]

Potential CON: Could increase the artifact size for certain (a lot of?) crates. Rn it's unclear to me how bad the fallout would be. Needs to be investigated.

[fmease]

Blocking this on #149019.

[fmease]

It should be as simple as only discarding positive impls if the ADT is a "boring" struct where "interesting" might be as simple as ty_param_count > 0 or more complex (TBD).

SATI is general enough to handle Sized even though it's not an auto trait / even though Sized cares about struct tails only while auto traits care about all constituent types. That's because we're dealing with well-formed types here. So for a struct all fields except the last one have to be Sized, otherwise compilation would've failed earlier, and thus SATI shouldn't generate "superfluous" or "misleading" predicates for them. Effectively, it should only be able to create "interesting" predicates for the very last field. For enums & unions, all fields have be Sized and we can assume that in rustdoc, meaning we shouldn't generate any "extra" predicates for these ADT kinds at all!