Discourage impl of trait not in your crate for type not in your crate

[dtolnay]

Such impls can lead to surprising and hard to debug inference failures in totally unrelated code. For example:

fn main() {
    let bools = vec![true, false];
    let stuff = Vec::new();
    assert_ne!(bools, stuff);
}

But adding this impl makes the inference break. The impl is PartialEq (not in our crate) for bool (not in our crate).

struct S;
impl PartialEq<S> for bool {
    fn eq(&self, other: &S) -> bool {
        unimplemented!()
    }
}

Real instance of this causing trouble: diesel-rs/diesel@1f3d7b1

[Ixrec]

This seems like it's harmful primarily when the trait and type are both from std. Maybe we could add a clippy lint against the general pattern of impl<...> [std trait]<...> for [std type]<...>?