Tracking Issue for `LintDiagnostic` migration

[JonathanBrouwer]

This is a tracking issue for the LintDiagnostic migration that @GuillaumeGomez and I are doing.
Mostly meant to be a place we can put ideas and keep track of what is left to do.

Implementation history

• The original PR that was split up: Remove LintDiagnostic #152811
• Start migration for LintDiagnostic items by adding API and migrating LinkerOutput lint #152933
• Migration of LintDiagnostic - part 2 #153016
• Migration of LintDiagnostic - part 3 #153051
• Migration of LintDiagnostic - part 4 #153091
• Migration of LintDiagnostic - part 5 #153152
• Migration of LintDiagnostic - part 6 #153352
• Migration of LintDiagnostic - part 7 #153401
• Cleanup unused diagnostic emission methods #153452
• Cleanup unused diagnostic emission methods - part 2 #153509
• Remove TyCtxt::node_span_lint method #153621
• Rename rustc_middle::lint::diag_lint_level into lint_level #153787
• Rename opt_span_diag_lint into opt_span_lint and remove emit_diag_lint #153820

Removing BuiltinLintDiag

• Start migrating DecorateDiagCompat::Builtin items to DecorateDiagCompat::Dynamic #154120
• Remove more BuiltinLintDiag variants #154224
• Remove more BuiltinLintDiag variants - part 2 #154303
• Remove more BuiltinLintDiag variants - part 3 #154336
• Remove more BuiltinLintDiag variants - part 4 #154376
• Set up API to make it possible to pass closures instead of AttributeLint #154432
• Remove BuiltinLintDiag #154912

TODO list

• See if we can remove the span argument from opt_span_diag_lint Migration of LintDiagnostic - part 4 #153091 (comment)
  • Possible but because of lint_level, we will need to update Diagnostic::into_diag so that it looks like:

    fn into_diag<F: FnOnce(Span) -> bool>(self, dcx: DiagCtxtHandle<'a>, level: Level) -> Option<Diag<'a, G>>

    That will allow to return early on and in particular skip Ty translation into diagnostic message, which would call DiagCtxt::set_must_produce_diag, forcing the Diag to be emitted. So before it's even possible, we need to ensure that the Diag will actually be emitted, and if not, we return early with None.
• See if we can get rid of BuiltinLintDiag in favour of a boxed diagnostic of some kind
• See if we can undo the perf regression in decorate_builtin_lint, maybe by using boxed diagnostics everywhere