indexing with T extends keyof <Large Union> causes OOM if keyof has a number

[weswigham]

Ref #23977 and discussion on #24137

Code

Isolated repro (needs dom lib for ElementTagNameMap and associated types):

// @strict: true
// Modified repro from #23977

declare global {
    interface ElementTagNameMap {
        [index: number]: HTMLElement
    }

    interface HTMLElement {
        [index: number]: HTMLElement;
    }
}

export function assertIsElement(node: Node | null): node is Element {
    let nodeType = node === null ? null : node.nodeType;
    return nodeType === 1;
}
  
export function assertNodeTagName<
    T extends keyof ElementTagNameMap,
    U extends ElementTagNameMap[T]>(node: Node | null, tagName: T): node is U {
    if (assertIsElement(node)) {
        const nodeTagName = node.tagName.toLowerCase();
         return nodeTagName === tagName;
    }
    return false;
}
  
export function assertNodeProperty<
    T extends keyof ElementTagNameMap,
    P extends keyof ElementTagNameMap[T],
    V extends HTMLElementTagNameMap[T][P]>(node: Node | null, tagName: T, prop: P, value: V) {
    if (assertNodeTagName(node, tagName)) {
        node[prop];
    }
}

Expected behavior:
No crash.

Actual behavior:
OOM crash after a very long delay.

Preliminary investigation shows the OOM occurs when we try to get the distributed form of keyof <Large Union> - namely the process is defined recursively, so when the union has ~170 members, we create 170 intermediate types (flattening one member at a time, while also wasting time recalculating type flags and other bits we will never need) before getting the final, distributed result. Now, picture doing this on every relation comparison operation for P (this distribution is uncached), and multiplying it for V in assertNodeProperty - the complexity is huge. Our fast path doesn't fix this, since the numeric index signatures cause us to bail on the fast path as it currently is.

[mattmccutchen]

I ran into this while attempting to write real-world code (I can provide the example if it's of interest). I have a fix up at #25728.