Regression when using `#[expect(unused)]` more than once in an impl block

[chinedufn]

At some point between rustc nightly-2023-03-05 (works) and nightly-2023-07-01 (does not work) the following stopped compiling:

#![deny(warnings)]

struct X;
impl X {
    #[expect(unused)]
    fn unused();

    // This is failing with `warning: this lint expectation is unfulfilled`.
    // Only the first `#[expect(unused)]` in an impl block works as expected.
    #[expect(unused)]
    fn unused2();
}

fn main() {
    let _ = X;
}

RUSTFLAGS="-D warnings" cargo check

I have not bisected for the exact commit that broke this.

I have only tried to reproduce this error using #[expect(unused)]. It's possible that this is also broken for other lints.

PR: #114417

[fmease]

Without properly double-checking this, I'm certain that this regressed in #110277 (merged April 14) which condensed multiple dead_code diagnostics triggered by methods into a single one for readability meddling with the affairs of lint_reasons.

I don't know how lint_reasons is implemented but if it only considers the first primary span of a (buffered) lint diagnostic, then one fix will be changing the relevant part of the code to consider all primary spans.

@rustbot label T-compiler A-lint F-lint_reasons