This solves a subtle bug with `sticking` when `use_model_load_balancing`
is used.

The problem was that for `Web/Sidekiq`:

1. If prior request would have `use_model_load_balancing=false` it would
   see load balancers to be `main/main`. This would result in sticking context
   to include only `main`.

2. If next request would evaluate `use_model_load_balancing=true` it would
   see load balancers for `main/ci`, but would only for wal locations see `main=lag`.
   As a result we would not check replication lag against `ci`.

3. This result in a situation that if `ci` is way behind `main`, the 2.
   would not understand the state of `ci`. Thus would consider `ci` to be up-to date
   and stick to `main`.

This commit fixes that:

1. If `ci:` is configured we always store the `ci:` in replication lag. This makes
   us the `wal_locations` (and `RackMiddleware` sticking context) to always include
   `main+ci` in all cases.

2. This results would behave (as for number of requests): we are reading primary / replica
   location from all databases.

3. Since we always have `ci` even if `use_model_load_balancing=false` we
   can properly evaluate state of replicas.