Dedup elaborated predicates with const generic parameter in AutoTrait

[megakorre]

Fixes #107715

Explanation

The param_env passed in to evaluate_predicates in the added test case has the constant N substituted for the value of the referencing constant 1. But the call to elaborate_predicates produces non normalized obligations (without the constant being substituted).

This means we will call SelectionContext.select with a param_env with multiple Predicates that will relate and cause a ambiguity. And eventually run in to a panic!.

Other normalization workarounds already happens in the AutoTraitFinder
in the add_user_pred method to work around lifetime differences.
But that method is called before elaborate_predicates which is what ends up adding the non normalized predicate.

This PR adds extra de-duplication with normalization that ends up removing the redundant predicate that causes the error here.

The first place I tried to put this de-duplication was in the Elaborator (this).
But the normalize utility I'm calling is not available from rustc_infer and the comments in AutoTraitFinder seems to suggest that the need to normalize the predicates like this is unique to it's use.

Existing Regression

running cargo-rustc-bisect finds that the test example was working before d49e7e7 (PR: #103279).
The removal of the code here makes the test pass. Not because the predicates then gets normalized the same. The param_env will still have extra predicates in it. But without the extra eval in the relate_consts code SelectionContext.select will only pick on off the 2 bounds and there will not be a ambiguity.
Making sure there are not multiple duplicates in the param_env seems more in line with the rest of AutoTraitFinder.

generic_const_exprs

With generic_const_exprs turned on the example works before and after this PR. The predicates get normalized to leaving the const un-evaluated in all cases.
So I left a comment that this de-duping can probably be removed when that feature is stable. this code has a similar comment.

Questions:

• I could add a extra test case that generates docs for the same test case but with generic_const_exprs enabled. That functionality was not broken before this so maybe its not needed? 🤷🏻 But the functionality of this is different in that case

[rustbot]

r? @nagisa

(rustbot has picked a reviewer for you, use r? to override)

[bors]

☔ The latest upstream changes (presumably #108250) made this pull request unmergeable. Please resolve the merge conflicts.

[nagisa]

r? @lcnr

[lcnr]

r? @BoxyUwU

[apiraino]

I'll play with the ball, too

r? compiler

[oli-obk]

@rustbot blocked on #108503

[fmease]

@megakorre Could you update the snippet issue: #107715 in the PR description to Fixes #107715 so GitHub can prominently link to this PR from the issue (I almost missed that there is a fix) and autoclose it when this PR gets merged? That would be great, thanks!

[Dylan-DPC]

Closing this as the issue this is referencing is already closed/fixed.