Canonicalize all nested goals together at once when making response

[compiler-errors]

Since we canonicalize each added goal separately, any time we add two goals that should be linked by an inference variable local to our evaluation, they end up unlinked in the nested goals of a inspect candidate. For example,

impl<T, U> Foo for T
where
    T: Constrain<Output = U>,
    U: Bar,
{
}

impl Constrain for () {
    type Output = ();
}

when proving (): Foo, we expect two nested goals: <() as Constrain>::Output = ?0 and ?0: Bar. However, we end up actually instantiating the second goal as ?1: Bar, leaving it totally unconstrained. This leads to anomalous states, such as a candidate with Certainty::Yes but with Certainty::Maybe in its nested goals❗

This PR aims to fix this problem by canonicalizing all of the nested goals all together at once, which means that we'll canonicalize any duplicated instances of a variable as a single variable.

r? lcnr

[rustbot]

Some changes occurred to the core trait solver

cc @rust-lang/initiative-trait-system-refactor

[compiler-errors]

🤦 needed to collect all goals

[compiler-errors]

Chose to use a generator here but I could very well write this using a flat map or something, if you'd prefer. Or just return a vec.

[lcnr]

I think we should instead add the fresh vars to the proof tree in fresh_args_for_item, we already do this when creating unconstrained vars for structurally_normalize

[compiler-errors]

oh lol that's so much easier to do