Fix a race condition in add_tunnel_endpoint()

Fix a race condition in add_tunnel_endpoint()

If there are multiple OVS agents concurrently executing
'tunnel_sync' RPC call a race condition can occur
leading to insertion of two different TunnelEndpoint
entries having the same 'id' value.

Unfortunately, we can not rely on:
  - @lockutils.synchronized(), because a Neutron installation can use
    more than one API node
  - with_lockmode('update'), because it works differently in PostgreSQL
    comparing to MySQL and doesn't guarantee that no new rows have been
    added to the table since the select query was issued. Please take a
    look at http://www.postgresql.org/files/developer/concurrency.pdf for
    more details.

The proposed fix:
  - ensures there is a unique constraint set for 'id' column
  - wraps creation of a new TunnelEndpoint entry into a
    repeatedly executed transactional block (so even if a concurrent
    DB transaction has been flushed or commited earlier than this one
    we can handle an integrity error and try again, in spite of the
    specified transactions isolation level value)

Fixes bug 1167916

Change-Id: I62dc729d595f090436199d5e1b6b98a884ead7a5