[FIX] This complements commit 3511,

revision-id: odo@openerp.com-20110714105552-9tgofrjtdgjmgc4b.
Each OpenERP cursor is mapped to a single psycopg2 connexion.
When a cursor is closed, the connexion is pushed back to a
pool and reused later. Now that the 'snapshot isolation' level
is used, the fact we didn't properly commit/rollback a
transaction appears: some 'concurrent' update showed up.
The fix is simple: whenever a cursor is closed, we rollback
any pending operation (which is the expected behavior).
(Furthermore, the connexion is explicitely closed when the
connexion is pushed back but not kept in the pool.)

bzr revid: vmt@openerp.com-20110913143444-s49r7r2h6m00p5s3

openerp/sql_db.py

@@ -294,6 +294,9 @@ class Cursor(object):
         del self._obj
         self.__closed = True
 
+        # Clean the underlying connection.
+        self._cnx.rollback()
+
         if leak:
             self._cnx.leaked = True
         else:
@@ -430,6 +433,7 @@ class ConnectionPool(object):
                     self._debug('Put connection to %r in pool', cnx.dsn)
                 else:
                     self._debug('Forgot connection to %r', cnx.dsn)
+                    cnx.close()
                 break
         else:
             raise PoolError('This connection does not below to the pool')