The current implementation of `optimize_npartitions` of the Dask backend
queries information about the workers from the Dask client object. The
information is stored in the `client.scheduler_info()` return value
which is a dictionary that can have the key `workers`.

Supposedly, when this key exists it means the Dask client has the needed
information. This is not always true. In certain scenarios, for example
when waiting for a batch system to return the available workers to the
dask client, the `workers` key will be present but its value will be an
empty dictionary. This is because the scheduler doesn't already know
which nodes of the cluster will become workers (this can be mitigated by
calling the `client.wait_for_workers` function beforehand).

This commit makes the check a bit stronger, getting the value of the
dictionary key `workers` and then checking if that value actually
contains something.