feat(compiler): Bring back the Map<K, V> type

[siketyan]

Related Issue: #2842

I was wondering how to emit propertyNames keyword of a object schema from TypeSpec. Then I found #2842. After looked up #2842 (comment), #571, and #1537, I noticed the previously there was the Map<K, V> type. (I couldn't catch exactly why it has been removed)

Because of lacking the key type of Record<T>, we cannot inject some custom types as the keys of the object. As @timotheeguerin suggested, bringing back the Map type should be the start point for adding propertyNames support.

As the current main branch of TypeSpec doesn't support OpenAPI 3.1 yet, I postponed to add emitting support of propertyNames keyword. Once this pull request got merged, I will prepare it, targeting the openapi3.1 branch.

[ArcturusZhang]

does this mean the key here could be an enum, or scalar or union of enums and/or scalars?
How should we understand the Union here?
would a union of models be OK here?
Also I did not see any test cases in this PR when key is an enum or union, therefore I am really curious what the union here stands for

[siketyan]

I added Union here for inline enum-like types, such as "dog" | "cat". AFAIK "dog" will be a Scalar type but "dog" | "cat" sill be a Union type.

Union of models won't be ok here, at least the key type must be assignable to string for Maps, by the extends keyword. However the ModelIndexer type cannot be constrained to such a limitation.

Please let me know if there is a better approach! 🙏 I will add missing test cases for enums though.

[ArcturusZhang]

so the union here actually means union of literals of enums and scalars.
if the compiler could prevent all those unexpected usage scenarios, this is actually fine

[siketyan]

Indeed. The compiler can check the key type for Map types, but IIRC the compiler currently doesn't check argument types for @indexer decorator as it's an intrinstics decorator without any declarations.