Revert "Cleanup markdown span handling"

[jyn514]

Reverts #80244. This caused a diagnostic regression, originally it was:

warning: unresolved link to `std::process::Comman`
 --> link.rs:3:10
  |
3 | //! [a]: std::process::Comman
  |          ^^^^^^^^^^^^^^^^^^^^ no item named `Comman` in module `process`
  |
  = note: `#[warn(broken_intra_doc_links)]` on by default

but after that PR rustdoc now displays

warning: unresolved link to `std::process::Comman`
 --> link.rs:1:14
  |
1 | //! Links to [a] [link][a]
  |              ^^^ no item named `Comman` in module `process`
  |
  = note: `#[warn(broken_intra_doc_links)]` on by default

which IMO is much less clear.

cc @bugadani, thanks for catching this in #77859.
r? @GuillaumeGomez

[rust-highfive]

Some changes occurred in intra-doc-links.

cc @jyn514

[GuillaumeGomez]

Could you also add a UI test to prevent a future regression too please?

[jyn514]

@GuillaumeGomez done, good idea.

[GuillaumeGomez]

Thanks! A bit sad for this change to be reverted but I guess it's going back to better jump. 😉 r=me once CI pass

[jyn514]

@bors r=GuillaumeGomez

[bors]

📌 Commit 0f25712 has been approved by GuillaumeGomez

[bugadani]

Some notes:

• CowStr is Borrowed only as long as the original link destination doesn't have to be transformed. This covers most of the intra-doc links, except for anchors that contain e.g. html entities.
• the original locate can handle the Borrowed case well enough, even though it's unsafe and ugly. But, since the borrowed string is part of the original source, it can be simplified. I suspect we can devise test cases where even the original code provides error messages pointing to the place of use instead of the definition. I'm not sure we can work around that, though, short of manually extracting the link label and looking up the definition in the md source.
• pulldown_cmark doesn't provide any information about reference-style link definitions. This is unfortunate, especially since Footnote defs are otherwise emitted as events.
• The only case where we have to care about all this (meaning the reason for reverting the original PR, and the points in this comment) is shortcut links and reference-style links, e.g. where the link definition is separate from the place of use.

[jyn514]

This fixes a diagnostic regression that was introduced in 1.50, but the fix was only merged in 1.51.

@rust-lang/rustdoc how do you feel about backporting this? The change is pretty small and only reverts an earlier change.

[GuillaumeGomez]

I agree with the backport.

[apiraino]

beta backport discussed and approved during today's meeting