forward extremities are leaked when their successors are persisted as outliers and later de-outliered

This issue has been migrated from [#6607](https://github.com/matrix-org/synapse/issues/6607).

---

Consider a sequence of events

```
A <- B <- C <- D
```

Starting with A, it is the only forward extremity in the DAG.

Then, suppose we receive D, and for whatever reason, the missing-events algorithm ends up retrieving B and C, and B is persisted as an outlier. We now have two forward extremities: A and D.

[B can be an outlier even though we have all its predecessors and successors: it means that we haven't resolved the state at that point in the graph.]

Later, we might de-outlier B due to a backfill request. When this happens, A remains a forward extremity.