Tag the inserted lifetime bounds with the span of the other bounds

[notriddle]

Fixes rust-lang/rust#93828 (which is really a bug in async_trait, not rustc).

Before:

help: consider further restricting this bound
   |
13 |     async fn publish<T + std::marker::Send: IntoUrl>(&self, url: T) -> String {
   |                        +++++++++++++++++++

After:

help: consider further restricting this bound
   |
13 |     async fn publish<T: IntoUrl + std::marker::Send>(&self, url: T) -> String {
   |                                 +++++++++++++++++++

[dtolnay]

I am hesitant to accept this because I don't think the change in the PR makes things consistently better. Obviously the ui test you gave, with a single trait bound which is a single identifier, gets better with this change. But a bunch of other errors get disproportionately worse:

Before:

   |
15 |     async fn publish<T + std::marker::Send: Trait<u8>>(&self, url: T) {}
   |                        +++++++++++++++++++

After, stable rustc:

   |
15 |     async fn publish<T: Trait + std::marker::Send<u8>>(&self, url: T) {}
   |                               +++++++++++++++++++

After, nightly rustc:

   |
15 |     async fn publish<T: Trait<u8>> + std::marker::Send(&self, url: T) {}
   |                                    +++++++++++++++++++

Also as you wrote in the comment in the ui test, fundamentally this is something that can't be accomplished by hacking around with spans. "It ought to insert the colon, but getting it to do that would require a way to tell rustc that the bounds don't actually exist in the source code, and it needs to insert them."

So I would prefer to see a fix in rustc instead. It could place the + Send suggestion at the max of the spans of all of the type parameter's trait bounds, instead of blindly after the span of the last trait bound. That way macro-generated trait bounds which are spanned with type parameter's span (async-trait is definitely not the only macro that does this) don't wreck the suggestion placement.