dispatch: adjust ui.flush() timing to stabilize test-blackbox.t

Without this change, dispatch.dispatch() could return before flushing all
stdio data. This means chg stdio would print data after receiving the result
code.

--- tests/test-blackbox.t
+++ tests/test-blackbox.t.err
@@ -354,13 +354,13 @@
   > EOF
   $ hg log --debug
   removing $TESTTMP/gone/.hg
-  warning: cannot write to blackbox.log: $ENOENT$ (no-windows !)
   warning: cannot write to blackbox.log: $TESTTMP/gone/.hg/blackbox.log: $ENOTDIR$ (windows !)
   $ cd ..

 blackbox should disable itself if track is empty

   $ hg --config blackbox.track= init nothing_tracked
+  warning: cannot write to blackbox.log: $ENOENT$
   $ cd nothing_tracked
   $ cat >> .hg/hgrc << EOF
   > [blackbox]

mercurial/dispatch.py

@@ -323,7 +323,7 b' def dispatch(req):'
             ret = -1
         finally:
             duration = util.timer() - starttime
-            req.ui.flush()
+            req.ui.flush()  # record blocked times
             if req.ui.logblockedtimes:
                 req.ui._blockedtimes[b'command_duration'] = duration * 1000
                 req.ui.log(
@@ -346,6 +346,8 b' def dispatch(req):'
                 req._runexithandlers()
             except:  # exiting, so no re-raises
                 ret = ret or -1
+            # do flush again since ui.log() and exit handlers may write to ui
+            req.ui.flush()
         return ret