[compiler] Repro for missed case of global mutation

I realized a pattern of global mutations that we don't currently detect (both in the old and new inference models). If you hide the mutation inside a function returned from another function, we lose track of it:

```js
const f = () => () => {
  global.property = true;
};
f()();
```

Roughly speaking, we need to track that if the return value of `f` is mutated, that it should count as triggering some effects. Right now we encode the idea that a function specifically can have side effects if it is mutated, but other values don't have a way to represent this. I'm thinking that we change the shape of the `Create` effect a bit, and allow room for an optional "mutation effects" array. Then, InferMutationAliasingRanges can visit these effects like it does when trying to find transitive function effects.