Various fixes related to the `durable_sessions` feature:

- Add an option to execute read operations on the leader.
- `drop_generation` operation can be replayed multiple times by the replication layer, but it's not idempotent. This PR adds a workaround that avoids a crash when `drop_generation` doesn't succeed. In the future, however, we want to make `drop_generation` idempotent in a nicer way.
- Wrap storage layer events in a small structure containing the generation ID, to make sure events are handled by the same layout CBM & context that produced them.
- Fix crash when storage event arrives to the dropped generation (now removed `storage_layer:generation_at` function didn't handle the case of dropped generations).
- Implement `format_status` callback for several workers to minimize log spam
- Move the responsibility of `end_of_stream` detection to the layout CBM. Previously storage layer used a heuristic: old generations that return an empty batch won't produce more data. This was, obviously, incorrect: for example, bitfield-LTS layout MAY return empty batch while waiting for safe cutoff time.
- `reference` layout has been enabled in prod build. It could be useful for integration testing.
- Fix incorrect epoch calculation in `bitfield_lts:handle_event` callback that lead to missed safe cutoff time updates, and effectively, subscribers being unable to fetch messages until a fresh batch was published.