Introduce types.anything

[infinisil]

Motivation for this change

This introduces the new type types.anything, which implements a safe and reasonable merge strategy if you don't care about the type. Specifically:

• It recursively merges attribute sets together
  • Except if it's string-coercible (aka, packages and co.), then only one definition is allowed
• It doesn't concatenate list values together (since list concatenation doesn't always make sense)
• If it's an atomic value (not a set or a list), all definitions must be equal
• It applies mkIf's and co. for all values

This is in contrast with the often-used types.attrs and types.unspecified, which can be problematic because they don't adhere to above points. If both this PR and #97114 are merged, I can start deprecating those in favor of this one.

Note that the doc entry for this new type is intentionally kept small, because its behavior is very natural and doesn't need a special explanation.

Ping @rycee @roberth @Profpatsch

Things done

• Wrote a doc entry for the new type
• Wrote tests for the new type

[edolstra]

Hm, what's the use case for this? Presumably you always do care about the type of an option since you need to use it somewhere eventually.

[rycee]

@edolstra I don't think the anything type is very useful by itself but I would happily use, e.g., attrsOf anything instead of attrs or attrsOf unspecified. That is, when I know the type is an attribute set but not the specific content. I've encountered this mainly when defining settings-like options that gets serialized to some hierarchical configuration format. For example the programs.taskwarrier.config or programs.waybar.settings.*.modules options in Home Manager.

[infinisil]

Yeah exactly, the intention is for this type to replace types.attrs and types.unspecified, which are currently used when the type isn't known, but as mentioned they have some surprising behavior which you wouldn't expect.

[infinisil]

Changes since the last revision:

• I moved types.attrs to the bottom of the type list and added a warning about its problems
• Added an example for types.anything in the docs
• Fixed mergeEqualOption for when a single function is passed. Now it doesn't error out anymore. Also added a test case to types.anything for this

[Profpatsch]

LGTM

[roberth]

Thanks @infinisil !