We're going to cache the total open count separately, and then just perform
these counts on the list. We already do that to get the pagination information,
through Kaminari, and a future change will make Kaminari reuse the query results
from earlier in the request.

When an issuable's state changes, or one is created, we should clear the cache
counts for a user's assigned issuables, and also the project-wide caches for
this user type.

These cache a hash of counts by state, so the state isn't needed in the key
itself.

[ci skip]

This runs a slightly slower query to get the issue and MR counts in the
navigation, but caches by user type (can see all / none confidential issues) for
two minutes.

When we are filtering by a single project, and the current user has access to
see confidential issues on that project, we don't need to filter by
confidentiality at all - just as if the user were an admin.

The filter by confidentiality often picks a non-optimal query plan: for
instance, AND-ing the results of all issues in the project (a relatively small
set), and all issues in the states requested (a huge set), rather than just
starting small and winnowing further.