Rollup merge of #91387 - graydon:E0038-clarification, r=wesleywiser

Rollup merge of #91387 - graydon:E0038-clarification, r=wesleywiser

Clarify and tidy up explanation of E0038

I ran into E0038 (specifically the `Self:Sized` constraint on object-safety) the other day and it seemed to me that the explanations I found floating around the internet were a bit .. wrong. Like they didn't make sense. And then I went and checked the official explanation here and it didn't make sense either.

As far as I can tell (reading through the history of the RFCs), two totally different aspects of object-safety have got tangled up in much of the writing on the subject:
  - Object-safety related to "not even theoretically possible" issues. This includes things like "methods that take or return Self by value", which obviously will never work for an unsized type in a world with fixed-size stack frames (and it'd be an opaque type anyways, which, ugh). This sort of thing was originally decided method-by-method, with non-object-safe methods stripped from objects; but in [RFC 0255](https://rust-lang.github.io/rfcs/0255-object-safety.html) this sort of per-impossible-method reasoning was made into a per-trait safety property (with the escape hatch left in where users could mark methods `where Self:Sized` to have them stripped before the trait's object safety is considered).
  - Object-safety related to "totally possible but ergonomically a little awkward" issues. Specifically in a trait with `Trait:Sized`, there's no a priori reason why this constraint makes the trait impossible to make into an object -- imagine it had nothing but harmless `&self`-taking methods. No problem! Who cares if the Trait requires its implementing types to be sized? As far as I can tell reading the history here, in both RFC 0255 and then later in [RFC 0546](https://rust-lang.github.io/rfcs/0546-Self-not-sized-by-default.html) it seems that the motivation for making `Trait:Sized` be non-object-safe has _nothing to do_ with the impossibility of making objects out of such types, and everything to do with enabling "[a trait object SomeTrait to implement the trait SomeTrait](https://rust-lang.github.io/rfcs/0546-Self-not-sized-by-default.html#motivation)". That is, since `dyn Trait` is unsized, if `Trait:Sized` then you can never have the automatic (and reasonable) ergonomic implicit `impl Trait for dyn Trait`. And the authors of that RFC really wanted that automatic implicit implementation of `Trait` for `dyn Trait`. So they just defined `Trait:Sized` as non-object safe -- no `dyn Trait` can ever exist that the compiler can't synthesize such an impl for. Well enough!

However, I noticed in my reading-and-reconstruction that lots of documentation on the internet, including forum and Q&A site answers and (most worrying) the compiler explanation all kinda grasp at something like the first ("not theoretically possible") explanation, and fail to mention the second ("just an ergonomic constraint") explanation. So I figured I'd clean up the docs to clarify, maybe confuse the next person less (unless of course I'm misreading the history here and misunderstanding motives -- please let me know if so!)

While here I also did some cleanups:

  - Rewrote the preamble, trying to help the user get a little better oriented (I found the existing preamble a bit scattered).
  - Modernized notation (using `dyn Trait`)
  - Changed the section headings to all be written with the same logical sense: to all be written as "conditions that violate object safety" rather than a mix of that and the negated form "conditions that must not happen in order to ensure object safety".

I think there's a fair bit more to clean up in this doc -- the later sections get a bit rambly and I suspect there should be a completely separated-out section covering the `where Self:Sized` escape hatch for instructing the compiler to "do the old thing" and strip methods off traits when turning them into objects (it's a bit buried as a digression in the individual sub-error sections). But I did what I had time for now.