`AliasRelate` hides info in transitive cases

[lcnr]

if we have A == alias == B and A and B are partially inferred, AliasRelate is weaker than the current eager_inference_replacement.

Due to our new approach to generalization in rust-lang/rust#119106 this example now compiles in the new solver. While the underlying issue is still not fixed, it feels incredibly hard to trigger now.

use std::marker::PhantomData;

#[derive(Default)]
struct Foo<T, U>(PhantomData<(T, U)>);

trait Trait {
    type Assoc;
    
    fn to_assoc(self) -> Self::Assoc;
}

impl Trait for Foo<u32, i32> {
    type Assoc = Foo<u32, i32>;
    fn to_assoc(self) -> Self::Assoc {
        Foo(PhantomData)
    }
}
impl Trait for Foo<i32, u32> {
    type Assoc = Foo<i32, u32>;
    fn to_assoc(self) -> Self::Assoc {
        Foo(PhantomData)
    }
}

fn main() {
    let mut x: Foo<_, _> = Default::default();
    let mut assoc = x.to_assoc();
    assoc = Foo::<u32, _>(PhantomData);
    assoc = Foo::<_, i32>(PhantomData);
    x = assoc;
}

passes on stable, results in the following error with the new solver

error[E0282]: type annotations needed
  --> <source>:26:12
   |
26 |     let x: Foo<_, _> = Default::default();
   |            ^^^^^^^^^ cannot infer type for struct `Foo<_, _>`

error[E0283]: type annotations needed
  --> <source>:27:23
   |
27 |     let mut assoc = x.to_assoc();
   |                       ^^^^^^^^
   |
note: multiple `impl`s satisfying `Foo<_, _>: Trait` found
  --> <source>:12:1
   |
12 | impl Trait for Foo<u32, i32> {
   | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
...
18 | impl Trait for Foo<i32, u32> {
   | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
help: try using a fully qualified path to specify the expected types
   |
27 |     let mut assoc = <Foo<_, _> as Trait>::to_assoc(x);
   |                     +++++++++++++++++++++++++++++++ ~

The new solver has 3 obligations: AliasRelate(<Foo<_, _> as Trait>::Assoc, Foo<u32, _>) and AliasRelate(<Foo<_, _> as Trait>::Assoc, Foo<_, i32>) and then AliasRelate(<Foo<_, _> as Trait>::Assoc, Foo<_, _>) from the assignment x == assoc.

TODO: I expect us to keep shallow norm inside of the trait solver, so we need to add a test where we this happens while proving some goal