Only finalize mapped all from Go when map entry exists

There was still a double finalization race, detected by the panic introduced
by the previous commit.  If setErr() is called during client teardown, all
currently managed calls will be errored and finalized, and the calls map set
to nil.

Unfortunately, Client.Go was again finalizing the call even after setErr
released the calls mutex, because it had a another variable pointing to it.

To fix this race, look up the call entry in the map and only finalize and
delete it when it present.

Author: jrick

rpc.go

@@ -501,11 +501,16 @@ func (c *Client) Go(ctx context.Context, method string, result interface{}, done
 	if err != nil {
 		// This may be called concurrently with out() -> setErr(), so
 		// finalize the error under the client mutex to prevent double
-		// finalization.
+		// finalization.  Only call finalization if the map and entry
+		// still exists (as setErr() will nil the map it after client
+		// teardown), otherwise we will finalize the call again.
 		c.callMu.Lock()
-		delete(c.calls, id)
-		call.err = err
-		call.finalize()
+		call2, ok := c.calls[id]
+		if ok {
+			delete(c.calls, id)
+			call2.err = err
+			call2.finalize()
+		}
 		c.callMu.Unlock()
 	}
 	return call