Fix leak of skolemized types in HRTB

[nikomatsakis]

A subtle bug in the higher-ranked sub/lub/glb routines could cause incorrect results and/or leaked skolemized regions, something I've seen from time to time but never quite knew why. Finally tracked it down. Didn't open an issue but there are various tests that report wrong result and/or ICE in the PR.

r? @pnkfelix

[nikomatsakis]

Hmm, running this test to completion I see that it breaks some tests. (I think that those tests should not have been working in the first place, but what they are doing is fundamentally useful.) The tests are all cases where Clone must be implemented for fn(&int); the existing impl works only for types not containing bound regions. I think the only real fix to this is to have some kind of blanket impl that applies to all fn types, perhaps a meta-trait identifying fn pointers.

[reem]

I remember the blanket Clone impl being broken for fn(&int) a while ago, unless it was fixed and this is re-breaking it that seems unrelated.

[reem]

Also, is this related to the lack of an impl of Fn(&_,) for fn(&_)?

[pnkfelix]

nit: "algorith" => "algorithm"

[nikomatsakis]

Going to do more work in this area. I'll close this branch for now and re-open with more commits.