When using persistent stubs, it turns out that a stub produced by `ctx.restore()` holds a self-reference which can lead to a cyclic self-cancellation issue when tearing down the IoContext. We were deferring destruction of the worker, but not unlink(). We need to defer both.

If any externals are present, clearExternalsIfPresent() will execute a new query which causes the previous `query.changeCount()` to return the wrong value, because the change count comes from a SQLite API that reports from whatever query ran last.

Arguably, the `changeCount()` method is a bad interface but I've left fixing it as a TODO.

Fix two bugs affecting persistent stubs