librustc: Implement the unboxed closure trait hierarchy.

[pcwalton]

This makes FnMut require FnOnce and Fn require FnMut. Therefore,
this change breaks code that implements the FnMut and/or Fn traits
directly, without also implementing their dependencies. A simple
forwarding implementation that defines FnOnce in terms of FnMut
and/or Fn in terms of FnMut will suffice.

This does not affect code that simply uses the |&:|/|&mut:|/|:|
unboxed closure construction notation.

Part of RFC #44; needed to implement RFC #63.

Part of issue #12831.

[breaking-change]

r? @nikomatsakis

[nikomatsakis]

OK, so, I read through this patch and it all makes sense. However, I also remembered that what I originally had in mind was that we could (I think) alternatively create an impl for all FnMut in terms of Fn and so on (impl<A,R,T:Fn<A,R>> FnMut<A,R> for T) and hence have no actual extra work to do in the compiler. Of course this requires PR #17669 to work, since it is an example of conditional dispatch. Unless there is some reason this won't work that is escaping me at the moment.

[nikomatsakis]

@pcwalton the aliasing restrictions don't matter to trans...but by-value self is potentially different LLVM bitcode, yes.

[nikomatsakis]

@pcwalton I just drew up a quick sketch showing what I mean -- it seems to work, at least when all the impls are user-defined: 82d37da

Hat tip: @glaebhoerl suggested this approach first.

[alexcrichton]

ping @pcwalton, any progres here?

[nikomatsakis]

This should be fairly trivial to do now. I'm going to check it out.

[nikomatsakis]

Closing in favor of #18388