Support relations and inference between template literal types

[ahejlsberg]

With this PR we support relations and inference between template literal types. Specifically, when the target side in a type relation is a template literal type, the source side is now permitted to be a compatible (i.e. more specific) template literal type. Likewise, when inferring to a template literal target type, we now permit the source type to also be a template literal type.

Some examples of improved assignment relations:

declare let s1: `${number}-${number}-${number}`;
declare let s2: `1-2-3`;
declare let s3: `${number}-2-3`;
declare let s4: `1-${number}-3`;
declare let s5: `1-2-${number}`;
declare let s6: `${number}-2-${number}`;
s1 = s2;
s1 = s3;
s1 = s4;
s1 = s5;
s1 = s6;

With this PR all of the above assignments are permitted, where previously only the first assignment was permitted.

Some examples of inference between template literal types:

declare function foo1<V extends string>(arg: `*${V}*`): V;

function test<T extends string>(s: string, n: number, b: boolean, t: T) {
    let x1 = foo('*hello*');  // "hello"
    let x2 = foo('**hello**');  // "*hello*"
    let x3 = foo(`*${s}*` as const);  // string
    let x4 = foo(`*${n}*` as const);  // `${number}`
    let x5 = foo(`*${b}*` as const);  // "true" | "false"
    let x6 = foo(`*${t}*` as const);  // `${T}`
    let x7 = foo(`**${s}**` as const);  // `*${string}*`
}

Fixes #43060.
Fixes #43243.

[RyanCavanaugh]

Some sort of comment (an overview of the algorithm / what it's trying to do, or the name of it if this is a well-known one?) seems justified here

[DanielRosenwasser]

Yes, I think if you could just explain with examples too, that'd be helpful.

[ahejlsberg]

Ok, I'll add some comments.

[DanielRosenwasser]

Would it be useful to have a singleton emptyTemplateType?

[ahejlsberg]

Well, the single-element types array varies here, so we can't use a singleton. We could possibly have a singleton for the ["", ""] array, but it's hardly worth the effort as this code doesn't run that often.

[DanielRosenwasser]

I don't like calling these sourceTexts because it immediately signals to me "BUG" even though you're not actually grabbing out the source text. You're grabbing the normalized text content which is the correct thing to do.

[ahejlsberg]

Hmm, I don't think the current names are that bad. After all the properties are called texts and types in template literals and we want the parameters to reflect the relation.

[DanielRosenwasser]

Right, I think that even texts would be a better name - I may send a follow-up PR but I'm going to merge this for now.