explicit form `Fn<A,B>` and now should use `Fn(A) -> B` or
`Fn<A,Output=B>`, but in some cases we get duplicate error
reports. This is mildly annoying and arises because of the main error
and another error from the projection. Might be worth squashing those,
but seems like a separate problem.
Part of issue #16640. I am leaving this issue open to handle parsing of
higher-rank lifetimes in traits.
This change breaks code that used unboxed closures:
* Instead of `F:|&: int| -> int`, write `F:Fn(int) -> int`.
* Instead of `F:|&mut: int| -> int`, write `F:FnMut(int) -> int`.
* Instead of `F:|: int| -> int`, write `F:FnOnce(int) -> int`.
[breaking-change]
