Some API requests are handled by Grape directly without ever going
through `before` blocks. We can still log `remote_ip` (and hence
`client_id`) for those requests, as that doesn't depend on anything
application-specific.

We handle this by putting it in the logger itself as well as the
`before` block. This is because `client_id` is computed based on an
ApplicationContext instance, but this is backed by a Labkit::Context. A
Labkit::Context isn't aware of our models, and so `user` is only
inherited as a string, not an object - which means that `client_id`
could end up incorrectly using the IP rather than the user ID.