Outlives requirements are not implied in the return type of async fn, and generally for RPIT

[jackh726]

Split from #102682

The following doesn't compile:

trait MyTrait<T> {
    async fn foo(&self) -> &T;
}

it gives the following:

error[E0311]: the parameter type `T` may not live long enough
 --> <source>:2:5
  |
2 |     async fn foo(&self) -> &T;
  |     ^^^^^^^^^^^^^-^^^^^^^^^^^^
  |     |            |
  |     |            the parameter type `T` must be valid for the anonymous lifetime as defined here...
  |     ...so that the reference type `&T` does not outlive the data it points at
  |
help: consider adding an explicit lifetime bound
  |
2 |     async fn foo<'a>(&'a self) -> &'a T where T: 'a;
  |                 ++++  ++           ++   +++++++++++

of course, a normal function is fine. We just don't imply that T: '_ here because it lowers to impl Future<Output = &'_ T>

[compiler-errors]

I thought I mentioned in the issue that this has nothing to do with AFIT or RPITIT? This can be observed on regular RPIT and async fn:

async fn foo<'a, T>() -> &'a T { todo!() }
//~^ ERROR the parameter type `T` may not live long enough

[compiler-errors]

I've edited the PR title to reflect its more general nature. This is because we don't walk into RPIT's projections to gather outlives constraints, i.e. we don't deduce any implied outlives bounds from impl Future<Output = &'a T>.

I'm not certain this is fixable in general. Deducing outlives from the return type of a function is only really doable because the implementation of the body serves as the witness of the outlive's validity, but this is obviously not true for all associated types, just the ones that are "output"-like.