Better error message using byte string where only regular strings are allowed in attributes

[Havvy]

The error message shows the code that matters.

error: literal in `cfg` predicate value must be a string
 --> src/lib.rs:L:11
  |
L | #[cfg(X = b"a")]
  |    

It should probably suggest removing the b. Likewise if using single quotes, it should suggest using double quotes. Both should also be machine applicable. Raw strings are allowed, so nothing needs to be done with them.

[Havvy]

It might also be valid to give this an error code and explain it all there.

[Havvy]

@petrochenkov mentioned in #54929 that the error message should be done for all places a byte string might be erroneously used instead of an actual string literal. I agree, so I'm generalizing this issue out to that.

[zackmdavis]

Petrochenkov's technical-debt concern (which I agree with) also came up in #54683. I think the non-filling-the-compiler-with-garbage place to do this would have to be during parsing: perhaps parse_unsuffixed_lit could be generalized to take arguments specifying which literals are allowed (keeping parse_unsuffixed_lit as a less-general version that calls the generalized method)? (Struckthrough thought doesn't actually work because different attributes want to allow different kinds of literals and the parser doesn't and shouldn't know about the semantics of different attributes.)

[estebank]

Self quoting from #54929:

I feel that we can move a lot of these ad-hoc diagnostics further up in the parser by having a more complex version of Parser::eat(&mut self, TokenKind) that also takes some textual info for diagnostics, and also checks for "confusables"—not only single tokens that look alike, like ; and :, but also cases like this one (b""/"").