Fix rustdoc ICE caused by mishandling of ambiguity errors

[ShoyuVanilla]

Fixes #162557

pub trait Service<Request> {
    type Future;
}

pub trait ZebraService<Request>: Service<Request> {}

impl<MaybeVerify, Request> ZebraService<Request> for MaybeVerify where
    MaybeVerify: Service<Request, Future: 'static>
{
}

pub struct Verifier;

impl Service<()> for Verifier {
    type Future = &'static ();
}

In the above minimization of the issue, when we try to check whether the blanket impl can be applied to Verifier,

rust/src/librustdoc/clean/blanket_impl.rs

Lines 42 to 67 in a8a1e6f

	let args = infcx.fresh_args_for_item(DUMMY_SP, item_def_id);
	let impl_ty = ty.instantiate(tcx, args).skip_norm_wip();
	let param_env = ty::ParamEnv::empty();
	let impl_args = infcx.fresh_args_for_item(DUMMY_SP, impl_def_id);
	let impl_trait_ref = trait_ref.instantiate(tcx, impl_args).skip_norm_wip();
	// Require the type the impl is implemented on to match
	// our type, and ignore the impl if there was a mismatch.
	let Ok(eq_result) = infcx.at(&traits::ObligationCause::dummy(), param_env).eq(
	DefineOpaqueTypes::Yes,
	impl_trait_ref.self_ty(),
	impl_ty,
	) else {
	continue;
	};
	let InferOk { value: (), obligations } = eq_result;
	// FIXME(eddyb) ignoring `obligations` might cause false positives.
	drop(obligations);
	let clauses = tcx
	.clauses_of(impl_def_id)
	.instantiate(tcx, impl_args)
	.clauses
	.into_iter()
	.map(Unnormalized::skip_norm_wip)

We make fresh args for the where-clause and skip normalization for it.

So, we have <?MaybeVerify as Service<?Request>>::Future: 'static bound to check. But it immediately evaluated into ambiguity due to stalled on infer vars in the fast path as the obligation contains the infer vars.

But due #162182 we evaluated it directly in the solver without stalled_on and this time it succeeded because we can normalize the alias into a concrete type &'static () and it outlives the static region.

So, I think it's very iffy to call InferCtxt::evaluate_obligation on a non-rigid alias and we should eagerly normalize it in L67 from the above rustdoc code.
But it may break something in rustdoc as normalizations inside it is pretty messy in general 🫠 So, I guess in another PR with crater runs.

r? lcnr

[rustbot]

Some changes occurred to the core trait solver

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

These commits modify tests/rustdoc-json.
rustdoc-json is a public (but unstable) interface.

Please ensure that if you've changed the output:

• It's intentional.
• The FORMAT_VERSION in src/librustdoc-json-types is bumped if necessary.

cc @obi1kenobi