Added an "unresolved question" alluding to the potential importance of
better handling of covariance (and thus variance in general).

Namely, now officially propose that `Arena` be changed to `Arena<'a>`.

(As an aside, suggest that we *could* also provide the old API with an
`UnsafeArena` type; but that is not part of the proposal proper).

Props to huonw for pointing this out, since it forced me to think
through the cases where the near-parametricity of `T` matters.

Notably, it does not suffice to focus solely on the case where `v`
owns data of type `T`; one must also consider the scenario where the
destructor for `v` constructs a `T` on the fly.