Cut off inference for recursive mapped types

[sandersn]

Fixes #19951

Previously, when inferring to a self-referential (or otherwise recursive) homomorphic mapped type from a source type that also has recursive references, type inference would enter infinite recursion.

Now there is a more complex stack for mapped type inference. It mirrors the existing symbolStack but (1) includes the source type and (2) is passed through inferTypeForHomomorphicMappedType, which is actually called outside of inferTypes, and so restarts the symbolStack cache every time.

Note that it might be possible to avoid restarting inference inside inferTypeForHomomorphicMappedType and share the two caches by using a cache key. I only experimented briefly with this so far.

[weswigham]

function getInferenceStackKey(source: Type, target: Symbol) {
    return `${source.id},${getSymbolId(target)}`
}
// ...
const cacheKey = getInferenceStackKey(source, target.symbol);
if (mappedTypeCache && mappedTypeCache.has(cacheKey)) {
    return; // future: mappedTypeCache.get(cacheKey)
}
(mappedTypeCache || (mappedTypeCache = createMap())).set(cacheKey, true); // future: replace `true` with initialized candidate type which is filled out by the below `infer` call.
const inferredType = inferTypeForHomomorphicMappedType(source, <MappedType>target, mappedTypeCache);
mappedTypeCache.delete(cacheKey);

If the mappedTypeStack values are never needed other than for set membership checking, and the ordering is irrelevant other than for removal, then for efficiency (faster contains checks, no array allocation for each member) we should use a set and not a stack, right?

[sandersn]

Yes. I was going to do that next, but I decided to get @ahejlsberg's opinion about the general approach beforehand.