Auto merge of #116734 - Nadrieril:lint-per-column, r=cjgillot

Lint `non_exhaustive_omitted_patterns` by columns

This is a rework of the `non_exhaustive_omitted_patterns` lint to make it more consistent. The intent of the lint is to help consumers of `non_exhaustive` enums ensure they stay up-to-date with all upstream variants. This rewrite fixes two cases we didn't handle well before:

First, because of details of exhaustiveness checking, the following wouldn't lint `Enum::C` as missing:
```rust
match Some(x) {
    Some(Enum::A) => {}
    Some(Enum::B) => {}
    _ => {}
}
```

Second, because of the fundamental workings of exhaustiveness checking, the following would treat the `true` and `false` cases separately and thus lint about missing variants:
```rust
match (true, x) {
    (true, Enum::A) => {}
    (true, Enum::B) => {}
    (false, Enum::C) => {}
    _ => {}
}
```
Moreover, it would correctly not lint in the case where the pair is flipped, because of asymmetry in how exhaustiveness checking proceeds.

A drawback is that it no longer makes sense to set the lint level per-arm. This will silently break the lint for current users of it (but it's behind a feature gate so that's ok).

The new approach is now independent of the exhaustiveness algorithm; it's a separate pass that looks at patterns column by column. This is another of the motivations for this: I'm glad to move it out of the algorithm, it was akward there.

This PR is almost identical to rust-lang/rust#111651. cc `@eholk` who reviewed it at the time. Compared to then, I'm more confident this is the right approach.

Author: bors