rational: make sure Hash agrees with Eq

[cuviper]

We can't use a derived Hash when we have a manual Eq, because we
need to uphold the invariant a == b → h(a) == h(b). Since Eq
doesn't require them to be in reduced form, Hash also needs to be
normalized.

Fixes #310.

[cuviper]

Technically this is a breaking change, because the former derived Hash only required T: Hash, and here I've used T: Clone + Integer + Hash. However, every other operation on Ratio<T> requires Clone + Integer, and we can't fix the relationship to Eq with just T: Hash.

(I don't actually need Clone here, but I think it's better to be consistent.)

[cuviper]

There's not even a way to create a Ratio<T> without T: Clone + Integer except through deserialization. I think the only practical way this will break is for a generic function that requires only T: Hash but expects to use Ratio<T> as Hash from that. I think this is outweighed by the need to fix Hash related to Eq.

Thanks to @mbrubeck for reviewing this with me on IRC.

bors r+

[bors]

Build succeeded

• continuous-integration/travis-ci/push