etcd driver implement allocate_unique_key in a raceful manner

[Dima Kuznetsov]

The code that allocates unique keys for etcd is susceptible to a race where a number of execution contexts will get value 1.

When a unique key is generated it is retrieved, incremented, and stored using compare&swap operation. If a key is not present, 1 is stored and returned.

If several contexts perform the lookup at the same time, all will choose 1 and write it to the database.

[Omer Anson (omer-anson)]

This bug needs two side-by-side solutions:
1. The etcd driver should avoid race conditions on the first unique ID allocation
2. All unique keys should be allocated by df-db init and df-db upgrade (Where new models are added between versions).

[OpenStack Infra (hudson-openstack)]

Reviewed: https://review.openstack.org/507814
Committed: https://git.openstack.org/cgit/openstack/dragonflow/commit/?id=a9de80cb5d523dba3dcc1da3bbf135ecd1ed2d46
Submitter: Jenkins
Branch: master

commit a9de80cb5d523dba3dcc1da3bbf135ecd1ed2d46
Author: Lihi Wishnitzer <email address hidden>
Date: Wed Sep 27 13:20:55 2017 +0300

    Allocate unique key in etcd using "create"

    Create use atomic transaction in order to verify that
    the value wasn't exist before updating the DB.

    This should solve our race-condition. If two (or more)
    allocations will occure at the same time, after the first
    allocation will succeed, the others will fail.

    Change-Id: Id72c7c35415a68c1eebc05ed3126f59793d5c475
    Closes-bug: #1713496