`fn try_merge_responses` behavior

[lcnr]

whenever we have multiple candidates we use fn try_merge_responses to try and merge them. This function is supposed to be complete.

We currently merge responses if they are all equal or alternatively if one of them has no external constraints.

This is quite fragile because it breaks the moment there are any region constraints as these sometimes contain '0: '0 constraints. We should remove those. We currently only opportunistically resolve region vars during canonicalization but the filtering should happen before then 🤔 It also does not work for any normalizes-to goal as their whole point is to constrain the term.

I feel like we should replace this with a proper subset check instead.

[compiler-errors]

(cc rust-lang/rust#112875 (comment), where this is needed because of some trivial '0: '0 region that should be filtered during compute_external_query_constraints, but can't be currently.)

[lcnr]

This currently has some non-trivial interactions with the ParamEnv candidate incompleteness: https://rust.godbolt.org/z/bq7K3zox3

cc #12

trait Trait<'a> {
    type Assoc;
}

impl<'a, T> Trait<'a> for T {
    type Assoc = ();
}

fn impls<'a, T: Trait<'a>>() {}
fn projects<'a, T: Trait<'a, Assoc = ()>>() {} 

// normalizing requires `'a == 'static`, the trait bound does not.
fn foo<'a, T: Trait<'static, Assoc = ()>>() {
    impls::<'a, T>(); // ok, no constraints from impl, prefer that one
    projects::<'a, T>();
    // lifetime error, projection candidates always have inference constraints
    // 2 candidates -> prefer `param_env` candidate
}