We've observed a scenario where there's a very large capture with
hundreds of collections all using schema inference. Among all of the
collections, publications due to inferred schema updates are quite
frequent. Because each collection publication results in a subsequent
publication of the capture and materialization, it results in
extremely frequent publications of those tasks, which has two
important consequences:

- It can prevent user-initiated publications from completing
  successfully due to `PublicationSuperceded` or `ExpectPubIdNotMatched`
  errors.
- It can prevent the materialization from making progress, because every
  time it's activated it must re-start any in-progress reads of source
  journals. There may also be a data-plane bug that's affecting this,
  but the frequent publications are certainly a factor.

So we've decided to slow down the rate of controller-initiated
publications to try to alleviate those symptoms. This introduces two new
fields to the publication status: `last_publication_time` and
`min_publication_interval`. Controllers will now wait at least
`min_publication_interval` since the `last_publication_time` before
publishing a spec in response to publication of dependencies.

The default value of `min_publication_interval` is 30 minutes, which is
somewhat arbitrary, but alignes with the default `update_delay` of
materializations. Since this value is part of the controller status,
we can update it on a per-task basis.