tracing: add Instrumented::span_and_inner_pin_mut

[sepointon]

This is useful for implementing poll-type functions on foreign traits.

Closes #2804.

There's some potential for bikeshedding about names; I don't care one way or another, and so I used the name that the internal code is already using on the projection type.

Most use cases probably won't need &mut Span as opposed to &Span, but the mutable reference is actually natural and fully general here: the inner value really does need to be a pinned mutable reference (because that's what's used in poll-type functions), so the Instrumented value also needs to be behind a mutable reference, so there's never an opportunity to create a second reference to the span while the reference to the inner value is held.

There's potential for a span_and_inner_pin_ref method, which also already exists on the projection type, but I don't know what the use case for exposing that is as opposed to using the existing methods (since mutable references and exclusivity aren't in play).

[sepointon]

Should I also be changing tracing_futures::Instrumented in parallel? #3456 has at least one other person confused about the state of play here.

[sepointon]

I decided tracing_futures::Instrumented probably has a long-term future as a vehicle for trait impls unacceptable for core tracing, so it makes sense to add a parallel method to it.

I note that, with this PR, tracing_futures::Instrumented could be rewritten as a wrapper around tracing::Instrumented (since the whole point of this PR is to expose the necessary bits to implement poll-style foreign traits), which would remove the duplicated logic that's a divergence hazard and reduce the amount of unsafe code, but that's not something that has to happen in this PR and so it makes sense to defer as a follow-up.