1. Previously, we were not removing existing access levels before
   creating new ones. This is not a problem for EE, but _is_ for CE,
   since we restrict the number of access levels in CE to 1.

2. The correct approach is:

    CE -> delete all access levels before updating a protected branch
    EE -> delete developer access levels if "developers_can_{merge,push}" is switched off

3. The dispatch is performed by checking if a "length: 1" validation is
   present on the access levels or not.

4. Another source of problems was that we didn't put multiple queries in
   a transaction. If the `destroy_all` passes, but the `update` fails,
   we should have a rollback.

5. Modifying the API to provide users direct access to CRUD access
   levels will make things a lot simpler.

6. Create `create/update` services separately for this API, which
   perform the necessary data translation, before calling the regular
   `create/update` services. The translation code was getting too large
   for the API endpoint itself, so this move makes sense.