account for impl-restricted traits in coherence

[amirHdev]

the original change broke aho-corasick because of SliceIndex
made future upstream impls opt-in with #[rustc_coherence_future_impls] which VaArgSafe now use

https://rust-lang.zulipchat.com/#narrow/channel/144729-t-types/topic/Impl.20restriction.20sadness/with/626411732

[jhpratt]

Overall, this seems to be well-implemented. If it is accepted, there would need to be better diagnostics for this type of of situation. The name of the attribute itself is also not clear imo. I feel obliged to mention that I am not a compiler reviewer and have only acted in an advisory role reviewing restriction PRs.

But to repeat what I said on Zulip, I'm not sure this is the right way to go. It would implicitly mean that it wouldn't be permitted to remove the impl restriction anywhere the attribute is present.