Do not create dangling &T in Weak<T>::drop
Since at this point all strong pointers have been dropped, the wrapped `T` has also been dropped. As such, creating a `&T` to the dropped place is negligent at best (language UB at worst). Since we have `Layout::for_value_raw` now, use that instead of `Layout::for_value` to avoid creating the `&T`.
This does have implications for custom (potentially thin) DSTs, though much less severe than those discussed in #80407. Specifically, one of two things has to be true:
- It has to be possible to use a `*const T` to a dropped (potentially custom, potentially thin) unsized tailed object to determine the layout (size/align) of the object. This is what is currently implemented (though with `&T` instead of `&T`). The validity of reading some location after it has been dropped is an open question IIUC (https://github.com/rust-lang/unsafe-code-guidelines/issues/188) (except when the whole type is `Copy`, per `drop_in_place`'s docs).
In this design, custom DSTs would get a `*mut T` and use that to return layout, and must be able to do so while in the "zombie" (post-drop, pre-free) state.
- `RcBox`/`ArcInner` compute and store layout eagerly, so that they don't have to ask the type for its layout after dropping it.
Importantly, this is already true today, as you can construct `Rc<DST>`, create a `Weak<DST>`, and drop the `Rc` before the `Weak`. This PR is a strict improvement over the status quo, and the above question about potentially thin DSTs will need to be resolved by any custom DST proposal.
3d93dfdf6e