incremental compilation flags #[rustc_on_unimplemented] as an unused attribute

[pnkfelix]

Spawned off of #58633

The following code will compile successfully the first time you build it with -C incremental=dir, then will fail the second time (again with -C incremental=dir).

Code:

// This is a case where we should not see the lint, but we are seeing
// it if incremental compilation is enabled.

#![feature(on_unimplemented)]
#![deny(unused_attributes)]

#[rustc_on_unimplemented = "invalid"]
trait Index<Idx: ?Sized> {
    type Output: ?Sized;
    fn index(&self, index: Idx) -> &Self::Output;
}

#[rustc_on_unimplemented = "a usize is required to index into a slice"]
impl Index<usize> for [i32] {
    type Output = i32;
    fn index(&self, index: usize) -> &i32 {
        &self[index]
    }
}

fn main() {
    Index::<usize>::index(&[1, 2, 3] as &[i32], 2);
}

Demonstration:

% rustc +nightly -C incremental=dir on-impl-works.rs
% rustc +nightly -C incremental=dir on-impl-works.rs
error: unused attribute
 --> on-impl-works.rs:7:1
  |
7 | #[rustc_on_unimplemented = "invalid"]
  | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  |
note: lint level defined here
 --> on-impl-works.rs:5:9
  |
5 | #![deny(unused_attributes)]
  |         ^^^^^^^^^^^^^^^^^

error: unused attribute
  --> on-impl-works.rs:13:1
   |
13 | #[rustc_on_unimplemented = "a usize is required to index into a slice"]
   | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

error: aborting due to 2 previous errors

---

This may seem like a low priority item since it is using #![feature(on_unimplemented)]; but we use that feature in libcore itself, so this problem is causing issues for people who want to use incremental compilation when making their locally bootstrapped build, which defaults to -D warnings.

[pnkfelix]

Probably the easiest short-term answer is just do something analogous to 81a6437 but for on_unimplemented.

[pnkfelix]

For the long term, I am vaguely worried that this may be an instance of a class of bugs with the interaction between lints and incremental compilation that we should be tracking.

For example, see also #52387