fix: Remove some deep normalizations from infer

[ShoyuVanilla]

Fixes #20975

The issue comes from this assertion:

rust-analyzer/crates/hir-ty/src/next_solver/normalize.rs

Line 48 in 4bf516e

assert_eq!(coroutine_goals, vec![]);

The deeply normalization goals returns ambiguous errors with type unspecified integer literal because each integer types are all valid candidates.
Normaly, they should be fallbacked into i32 but that's done manually in the type inference, not from the trait solving.

This PR removes deep normalizations on Ty but there are still two places calling it with T: TypeFoldable with non Ty.
I tried to remove the deep normalization entirely from infer, this function.

rust-analyzer/crates/hir-ty/src/infer/unify.rs

Lines 288 to 299 in 4bf516e

	// FIXME: We should get rid of this method. We cannot deeply normalize during inference, only when finishing.
	// Inference should use shallow normalization (`try_structurally_resolve_type()`) only, when needed.
	pub(crate) fn normalize_associated_types_in<T>(&mut self, ty: T) -> T
	where
	T: TypeFoldable<DbInterner<'db>> + Clone,
	{
	let ty = self.resolve_vars_with_obligations(ty);
	self.infer_ctxt
	.at(&ObligationCause::new(), self.trait_env.env)
	.deeply_normalize(ty.clone())
	.unwrap_or(ty)
	}

But it caused few regressions especially upon const generic parameters so I left them for now. (Our const evaluation are quite wrong so might need be fixed from there)

I think this is not overlaps with #20974 but this can be waited and rebased upon it if so 😄

[ShoyuVanilla]

This ugly function was introduced by me (with some visitors back then) to fix some errors on non-resolved nested opaque types during mir lowering but it isn't needed anymore 😄

[ShoyuVanilla]

Some tests are seemingly regressed but I believe they are okay because they are errornous types

[ShoyuVanilla]

Well, actually this is mostly done by #20974 😅 I'll just add a test and some minor changes - though this one is minor in overall 😅 - in this PR once it has landed