pgsql: Kludge around clone race in Pike core.

As reported by Jeff Hungerford:

When there is:
- No connection at present.
- Two or more threads try to open an initial connection to the database
  simultaneously.

You can trigger this:

Attempting to clone an unfinished program
/usr/local/pike/8.0.175/lib/modules/Sql.pmod/Sql.pike:342:
    Sql.Sql()->create("localhost:5432","foo","bar","CENSORED",UNDEFINED)

This could be caused by a race in the Pike-core class-cloning code which
is not racefree when two threads try to instantiate the same object
at roughly the same instant.

This patch attempts a workaround to detect the race and loop until
it finishes without problems.  It would be better if this could be
fixed in the Pike-core.

lib/modules/Sql.pmod/pgsql.pike

@@ -217,7 +217,15 @@ protected void create(void|string host, void|string database,
 
   if(!_port)
     _port = PGSQL_DEFAULT_PORT;
-  .pgsql_util.register_backend();
+
+  // FIXME Looks like a bug in the cloning of an object
+  // If we do not loop here, and this function is called in two threads
+  // simultaneously, *and* the .pgsql_util class has not been instantiated
+  // yet, then Pike can throw an "Attempting to clone an unfinished program"
+  // error.
+
+  while(catch(.pgsql_util.register_backend()))	// Placate race in cloning
+    sleep(0);					// Yield
   _shortmux=Thread.Mutex();
   reconnect();
 }