union-proof the 'generic mapped type' check

[KiaraGrouwstra]

Fixes #15756.

This incidentally also uncovered one other case (Form from keyofAndIndexedAccess) where generic mapped types had been applied in an unsound way w.r.t. union keys. That simplified heuristic worked there for the basic case of non-union types, but without it inference ended up no longer strong enough there.
That appears a piece of collateral damage here, though I'm under the impression the altered check here wasn't wrong about such reduction being unsound there.

To elaborate:

class Form<T> {
    private childFormFactories: {[K in keyof T]: (v: T[K]) => Form<T[K]>}

    public set<K extends keyof T>(prop: K, value: T[K]) {
        this.childFormFactories[prop](value)
    }
}

Reduction makes for {[K in keyof T]: (v: T[K]) => Form<T[K]>}[K] -> (v: T[K]) => Form<T[K]>. If one substitutes some union A|B for K, that makes for (v: T[A|B]) => Form<T[A|B]>, distinct from the (I believe) expected result here of ((v: T[A]) => Form<T[A]>) | ((v: T[B]) => Form<T[B]>). That's why I believe this reduction here was technically unsound even though it would have held for the base case of non-unions.

[msftclas]

@tycho01,
Thanks for having already signed the Contribution License Agreement. Your agreement was validated by Microsoft. We will now review your pull request.
Thanks,
Microsoft Pull Request Bot

[gcnew]

I like your approach very much! Both the union test and the use of a synthetic generic function. 👍

[KiaraGrouwstra]

@gcnew thanks! wish I had a better idea on some of the other issues as well 😅, for the most part I'm just opening PRs before even knowing what I'm doing yet haha

[KiaraGrouwstra]

Closing, fixed by #18042!