parser.y: properly abstract over to-be-created strings

[Radvendii]

This fixes #14642 and adds tests to prevent regression.

In eab467e, I had missed cases like "${"string"}", where the Expr is already an ExprString. This PR provides a wrapper around std::variant that correctly abstracts over when we have a string and when we have another Expr.

• parser.y: correctly abstract over to-be-constructed ExprString
• tests: add tests for dynamic attribute in let and inherit

Motivation

Context

---

Add 👍 to pull requests you find important.

The Nix maintainer team uses a GitHub project board to schedule and track reviews.

[Ericson2314]

Looks pretty good to me, but needs a formatter run.

[xokdvium]

One minor comment. Also would like to see more tests. I can take over if that's fine with you @Radvendii

[Radvendii]

I added a few more tests (some results still surprised me. inherit ${"foo"}; is valid but inherit "${foo"}"; is not).

You're also welcome to add more of course.

One problem is that this is such cursed nix code that the formatter doesn't know what to do with it and CI is failing.

1. does this count as a bug in the formatter that I should send their way? (the code it's failing to format doesn't succeed to parse but for a reason other than the one the formatter is giving)
2. can we disable the formatter for specific files?

[xokdvium]

Added a bit more tests (example from @piegamesde) and clarified some comments, added monostate instead of default initialized empty std::string_view.

[flokli]

@Ericson2314 do you plan to tag a release containing this fix?

[Ericson2314]

@flokli I haven't been involved with point releases very much at all, but yes that sounds like a good idea to me. I will raise it with the others.