`config/recurring.yml`. Run them using the least busy dispatcher
(largest polling interval) or a new default dispatcher if none.

This also introduces some new options for the CLI, to allow for
`--dispatch_only` and `--work_only`, and to skip recurring tasks.

And also setting a description that we can use from Mission Control in
the future.

A Scheduler, instead of Dispatcher, that will be created only when there
are recurring tasks to run.

Calling `set` on the job class returns a `ConfiguredJob`, and calling
`perform_later` on a `ConfiguredJob` doesn't yield the job to a block
passed as parameter, so we can't check if the enqueuing was successful
or not. To work around this, the only way is to instantiate the job
ourselves with the arguments, then call `enqueue` on that, and rely on
the result, which would be what `job_class.perform_later` would yield
to the block.

To match worker's `queues` configuration.

No need to have a class, can specify this using just a command for which
we'd provide a default RecurringtJob class that will just eval the command.