Inconsistent handling of `#[coverage(..)]` attributes that are malformed or misplaced

[Zalathar]

(Note: The coverage attribute is currently unstable; see #84605.)

While writing tests for #[coverage(..)] for #84605, I encountered various inconsistencies in diagnostics for using the attribute with incorrect syntax or on inappropriate items.

For example:

• Bare #[coverage] is silently permitted in various situations, even though it should never be legal.
  • E.g. on modules, on impl blocks, on trait, and on impl Trait.
• Using name-value syntax #[coverage = "off"] results in different error messages than other kinds of incorrect syntax, and sometimes results in multiple errors for the same attribute.
  • The error message also claims that bare #[coverage] is permitted, which is incorrect.
• Using #[coverage(off)] or #[coverage(on)] on things other than functions/methods/closures sometimes results in an error, and sometimes only a warning.
  • It seems to be an error on inappropriate items, but only a warning on (non-closure) expressions, even though it would have no effect in either case.
• There is no diagnostic for attaching multiple otherwise legal coverage annotations to a single item, even though it's unclear which one would prevail.

[veera-sivarajan]

@rustbot label -needs-triage +A-diagnostics +D-inconsistent +D-confusing