cc @jacobtomlinson

So previously we would pass the address along with serialized Futures for two reasons:

1. To verify that the client we're connecting to is connected to the right scheduler (this comes up in some advanced multi-client/multi-scheduler situations)
2. To create a new client and connect it to a scheduler if no client is found locally

When we call get_client() from within a worker, we don't need to worry about 2 (we already have a fine connection to the scheduler) but we do still need to verify that this is the right cluster for us to be on. Using the address for this verification was error prone, because network names are odd.

Now we also pass along the scheduler ID string, and use that for verification instead. This makes us a bit more robust to bad addresses.

Just given this a test with this MRE.

On this branch it hangs for me at client.publish_dataset(df=df).

@quasiben in case you're interested.